openvibers/openvibe.codes Commits
Tasks with attempts, a live event stream, and the engine deciding what happens next (C2) (#162)
Changes
@@@ -20,9 +20,9 @@ One place to send a coding task to the agent that fits it — by what the task n| The harness | `openvibe-codes` ([bin/openvibe-codes.js](bin/openvibe-codes.js), [harness/](harness/)) | Runs on a person's own machine with their CLIs and keys; nothing goes through OpenVibe. See **Use the harness** below. || Home | `/` | What the harness is, the agents in the catalog, how a task finds its agent, Improve OpenVibe, and an honest table of what works today and what comes next. || Get started | `/start` | Install, the agents and what each needs, the commands, permission levels, hand-offs, the event format and using it from code. |--| Run on your machine | `/run`, `/run/tasks/:id` ([server/http/run.js](server/http/run.js)) | A signed-in person starts a coding task in any browser and their own machine runs it: an OpenVibe Run job (class `code`, artifact `[email protected]`, `trust: user-owned`) that Codes starts for them in their project (project delegation: `X-OV-Subject`, `X-OV-Project`; [server/domain/run-client.js](server/domain/run-client.js)). Run places it only on a machine of theirs (added on openvibe.bot/machines), whose OpenVibe Node runs `openvibe-codes job` ([harness/job.js](harness/job.js)) in its job sandbox: clone, run the agents with hand-offs, hand back the summary, the changed files and the patch, and, when asked and the machine holds `GITHUB_TOKEN`, push a new branch (never the one it started from) and open a pull request. The page streams the job's lines (Run's job stream through Codes) and keeps the result (`run_tasks`, migration 0003). Each task is its owner's; account export and deletion cover them (ADR-033, [server/account-data.js](server/account-data.js)). The task page shows the agents' **session** with openvibe-shared/agent-session.js: steps grouped, each step's command, change and output (live while it runs), helpers nested, hand-offs and the plan; after the end, the session from `codes_session_lines` (migration 0007): the lines are kept as the stream arrives, back-filled from the machine's final log, and replayed at `GET /run/tasks/:id/session.json?after=` (the owner, or a service with `codes.task.read` and `X-OV-Subject`); tasks from before keep the old `log`. Run's `run.job.*` events finalize a task at `POST /internal/events` ([server/events-consumer.js](server/events-consumer.js), no polling but a 10-minute fallback), and every state change is published as `codes.task.queued|running|succeeded|failed|cancelled` from an outbox ([server/events/outbox.js](server/events/outbox.js), contracts 0.232.0). |++| Run on your machine | `/run`, `/run/tasks/:id` ([server/http/run.js](server/http/run.js)) | A signed-in person starts a coding task in any browser and their own machine runs it: an OpenVibe Run job (class `code`, artifact `[email protected]`, `trust: user-owned`) that Codes starts for them in their project (project delegation: `X-OV-Subject`, `X-OV-Project`; [server/domain/run-client.js](server/domain/run-client.js)). Run places it only on a machine of theirs (added on openvibe.bot/machines), whose OpenVibe Node runs `openvibe-codes job` ([harness/job.js](harness/job.js)) in its job sandbox: clone, run the agents with hand-offs, hand back the summary, the changed files and the patch, and, when asked and the machine holds `GITHUB_TOKEN`, push a new branch (never the one it started from) and open a pull request. The page streams the job's lines (Run's job stream through Codes) and keeps the result (`run_tasks`, migration 0003). Each task is its owner's; account export and deletion cover them (ADR-033, [server/account-data.js](server/account-data.js)). The task page shows the **timeline of attempts** above the agents' **session** (each attempt with its why: "Codex hit its limit, so Command Code took over with what it had done") and the session with openvibe-shared/agent-session.js: steps grouped, each step's command, change and output (live while it runs), helpers nested, hand-offs and the plan; after the end, the session from `codes_session_lines` (migration 0007): the lines are kept as the stream arrives, back-filled from the machine's final log, and replayed at `GET /run/tasks/:id/session.json?after=` (the owner, or a service with `codes.task.read` and `X-OV-Subject`); tasks from before keep the old `log`. Run's `run.job.*` events finalize a task at `POST /internal/events` ([server/events-consumer.js](server/events-consumer.js), no polling but a 10-minute fallback), and every state change is published as `codes.task.queued|running|succeeded|failed|cancelled` from an outbox ([server/events/outbox.js](server/events/outbox.js), contracts 0.232.0). || The agents | `/harnesses` | Every harness in [`server/data/harness-offers.json`](server/data/harness-offers.json) (validated at boot as `platform.harness-offer@1`, each model as `platform.agent-offer@1`) with its provider, address kind, what it can do, price and limits, and its models with context limits and token prices. No local path is ever shown. |--| Task API (services) | `POST /api/v1/tasks`, `GET /api/v1/tasks/:id`, `POST /api/v1/tasks/:id/cancel` ([server/http/tasks-api.js](server/http/tasks-api.js)) | Contracts 0.162 `codes.task.start` / `read` / `cancel`. A first-party service (OpenVibe.Actor) holding the capability starts a coding task for a person named in `X-OV-Subject`, with a Network service token for audience `openvibe.codes` (`CODES_AUDIENCE`). The task is the same one `/run` starts ([server/domain/run-tasks.js](server/domain/run-tasks.js)): it runs on that person's own machine, in the project of their latest `/run` task (none yet → 409 `codes.no_project`). It shows on their `/run` page. `idempotency_key` answers a retry with the first task (200, `created: false`). A service reads and cancels only the tasks it started for that person (otherwise 404). The result the machine returned is passed on only as `codes.task@1` allows, each field checked and cut to its limit (migration 0005: `via`, `idempotency_key`). Calls count against the person's budget: starts and cancels 6 a minute, 60 an hour. |++| Task API (services) | `POST /api/v1/tasks`, `GET /api/v1/tasks/:id`, `GET /api/v1/tasks/:id/events`, `POST /api/v1/tasks/:id/continue`, `POST /api/v1/tasks/:id/stop`, `POST /api/v1/tasks/:id/cancel` ([server/http/tasks-api.js](server/http/tasks-api.js)) | Contracts 0.162 `codes.task.start` / `read` / `cancel`. A first-party service (OpenVibe.Actor) holding the capability starts a coding task for a person named in `X-OV-Subject`, with a Network service token for audience `openvibe.codes` (`CODES_AUDIENCE`). The task is the same one `/run` starts ([server/domain/run-tasks.js](server/domain/run-tasks.js)): it runs on that person's own machine, in the project of their latest `/run` task (none yet → 409 `codes.no_project`). It shows on their `/run` page. `idempotency_key` answers a retry with the first task (200, `created: false`). A service reads and cancels only the tasks it started for that person (otherwise 404); the owner needs no token for their own tasks. The result the machine returned is passed on only as `codes.task@1` allows, each field checked and cut to its limit (migration 0005: `via`, `idempotency_key`). Calls count against the person's budget: starts and cancels 6 a minute, 60 an hour. **A task's attempts** (stage C2, migration 0010 `codes_attempts`): every Run job is one attempt, so a continue, a switch to another agent and a wrap-up are attempts on one timeline, never new tasks. `GET /api/v1/tasks/:id` answers `codes.task@2`: the task's own fields at the top level, exactly as `codes.task@1` had them, and, added to them, `attempts[]` (the timeline) and `outcome` — `done`, `unfinished` (with what is left) or `stopped`. `GET /api/v1/tasks/:id/events` is the task's event stream (server-sent events): the session's lines in order with the attempt each belongs to, the attempts themselves (started, ended with why), a heartbeat every 15 s, and a resume with `Last-Event-ID`. `POST /api/v1/tasks/:id/continue` starts the next attempt now (with `agent` it switches), `POST /api/v1/tasks/:id/stop` stops the running attempt and nothing carries it on. || Agent API | `GET /api/v1/harnesses`, `POST /api/v1/harnesses/route` | Public, no key. The router takes `{ task, requirements? }` (`task` is `edit`, `review`, `long` or `resume`; `requirements.capabilities` adds hard constraints) and answers `{ selected, reasons, candidates }` from `openvibe-sdk/placement` — hard constraints first (a missing capability excludes an agent, and says so), then cost. RFC 9457 problems (`harness.invalid`, `harness.task_unknown`). || Improve OpenVibe | `/improve` | Every OpenVibe repository, from the pinned contracts' service manifests (the services that run, then the shared libraries), and the path from a change to a reviewed pull request. || Community | `/policy`, `/policy/<document>` | The code of conduct ([CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md)), the contribution guide ([CONTRIBUTING.md](CONTRIBUTING.md)), the [contributor ladder](docs/governance/contributor-ladder.md) and the [moderation policy](docs/governance/moderation.md), rendered from those files. All four are **drafts pending owner review**: marked at the top, shown with a notice, not indexed; deleting the draft line publishes one. They govern the repositories and the apps and releases on OpenVibe.Services alike. |@@@ -84,7 +84,7 @@ openvibe-codes resume <run id> "now add a test"- **OpenVibe.Network** — SSO (OAuth client `codes`, PKCE S256) and its JWKS.- **PostgreSQL 18 and Valkey 9** (OpenVibe.Host `roles/data/`, ADR-035): `openvibe-sdk/db`; Valkey holds the per-actor limit counters (optional).--- **openvibe-contracts v0.143.0**, **openvibe-sdk v0.67.0**, **openvibe-shared v4.23.0** (pinned tag tarballs).++- **openvibe-contracts v0.243.0**, **openvibe-sdk v0.73.0**, **openvibe-shared v4.48.0** (pinned tag tarballs).No path in Codes sends or accepts a shared loopback key (tested by grep and at runtime).@@@ -158,7 +158,7 @@ Reporting a vulnerability: [SECURITY.md](SECURITY.md).Part of the [OpenVibe network](https://openvibe.network). Built in the open by [OpenVibers](https://github.com/OpenVibers).<!-- versions:start -->--- openvibe-contracts: v0.232.0++- openvibe-contracts: v0.243.0- openvibe-sdk: v0.73.0- openvibe-shared: v4.48.0<!-- versions:end -->@@@ -7,11 +7,11 @@"runtime": "node 22, express 4, PostgreSQL 18 (openvibe-sdk/db, async) + Valkey 9","port": 4900,"domain": "openvibe.codes",-- "contracts": "openvibe-contracts v0.232.0",++ "contracts": "openvibe-contracts v0.243.0","sdk": "openvibe-sdk v0.73.0","shared": "openvibe-shared v4.48.0","packages": [-- "openvibe-contracts v0.232.0",++ "openvibe-contracts v0.243.0","openvibe-sdk v0.73.0","openvibe-shared v4.48.0"],@@@ -84,8 +84,13 @@ function parseArgs(raw, env = process.env) {} : null;if (push && push.branch && (!BRANCH.test(push.branch) || push.branch.startsWith('-'))) throw new Error('push.branch must be a branch name');if (push && push.branch && PROTECTED.test(push.branch)) throw new Error(`push.branch must be a new branch of the task's own, never ${push.branch}`);++ // The session a resume continues: the job's own input, or RESUME in the environment (a task carried on in the++ // same conversation). It is a harness's session id, never a path and never a credential.++ const resumeIn = a.resume == null || a.resume === '' ? env.RESUME : a.resume;++ const resume = typeof resumeIn === 'string' && /^[A-Za-z0-9._:-]{1,200}$/.test(resumeIn) ? resumeIn : null;return {repo: repo.toString(), ref, prompt, permission, push,++ ...(resume ? { resume } : {}),agent: typeof a.agent === 'string' && /^[a-z0-9-]{1,40}$/.test(a.agent) ? a.agent : null,model: typeof a.model === 'string' && a.model.length <= 100 ? a.model : null,};@@@ -129,7 +134,7 @@ const resultText = (c) => (typeof c === 'string' ? c : Array.isArray(c) ? c.map(function compact(e) {const cut = (s) => cutTo(s, MAX_TEXT);const parent = e.parent_tool_use_id ? { parent: String(e.parent_tool_use_id) } : {};-- if (e.type === 'system' && e.subtype === 'init') return { t: 'agent', agent: e.agent, model: e.model || null };++ if (e.type === 'system' && e.subtype === 'init') return { t: 'agent', agent: e.agent, model: e.model || null, session: e.session_id || null };if (e.type === 'system' && e.subtype === 'handoff') return { t: 'handoff', from: e.from, ...(e.to ? { to: e.to } : {}), reason: cut(e.reason) };if (e.type === 'system' && (e.subtype === 'error' || e.subtype === 'warning')) return { t: 'note', level: e.subtype, text: cut(e.text) };if (e.type === 'system' && e.subtype === 'tool_progress') return { t: 'tool_progress', id: String(e.tool_use_id || ''), chunk: clean(cutTo(e.text, MAX_OUTPUT)) };@@@ -266,6 +271,9 @@ async function runJob({ input, harness, emit, writeResult, env = process.env, tmconst run = harness.run({prompt: job.prompt, task: 'edit', agent: job.agent, model: job.model, cwd: work,permission: job.permission, handoff: !job.agent, maxAttempts: 3,++ // A task carried on in the same conversation resumes its session (RESUME in the environment, or the job's++ // own resume field).++ resume: job.resume || null,});let last = null;for await (const e of run) {nnew file mode 100644@@@ -0,0 +1,109 @@++-- phase: expand++-- Stage C2 (openvibe.codes v1): a task's attempts. A continue, a switch to another agent and a wrap-up are attempts on++-- one task's timeline, never new tasks with cryptic names (the portal's "-c1"/"-r2" confusion of 10-30). Every Run job++-- is one attempt: the attempt carries the job, the agent, the machine and the harness session a resume needs.++--++-- n the attempt's place on the timeline (1, 2, …)++-- state queued | running | done | failed | stopped | budget-stopped | interrupted++-- reason limit | auth | capacity | refused | error | budget | stopped | hung | context (why it ended)++-- why one sentence in plain words: why this attempt happened (for the timeline)++-- kind first | continue | switch | wrap-up | manual: how this attempt came to be++-- follow_up_of the attempt whose end started this one. It is unique, so one ending can start only one follow-up:++-- the fence against the portal's 10-30 bug, where two racing ticks carried one job on twice.++--++-- run_tasks grows the task's own verdict: outcome ('done' | 'unfinished' | 'stopped', NULL while it runs), outcome_note++-- (what is left, in plain words) and attempts_count. codes_session_lines grows attempt_n, so every line of the session++-- says which attempt it belongs to.++--++-- Every task that exists now becomes attempt 1 of itself: the timeline of an old task is not empty.++++CREATE TABLE IF NOT EXISTS codes_attempts (++ id text COLLATE "C" PRIMARY KEY, -- att_<ULID>++ task_id text COLLATE "C" NOT NULL, -- tsk_<ULID>++ n integer NOT NULL, -- 1, 2, … the attempt's place on the timeline++ agent text COLLATE "C", -- the harness (claude-code, codex, …) or 'auto'++ model text COLLATE "C",++ machine text COLLATE "C", -- the node's id, or 'sandbox'++ job_id text COLLATE "C", -- Run's job_<ULID> for this attempt++ session text COLLATE "C", -- the harness session id, for a resume++ state text COLLATE "C" NOT NULL, -- queued, running, done, failed, stopped, budget-stopped, interrupted++ reason text COLLATE "C", -- limit, auth, capacity, refused, error, budget, stopped, hung, context++ why text COLLATE "C", -- one sentence in plain words++ kind text COLLATE "C" NOT NULL DEFAULT 'first',-- first | continue | switch | wrap-up | manual++ cost_usd numeric(14,6),++ tokens bigint,++ steps integer,++ started_at timestamptz,++ finished_at timestamptz,++ created_at timestamptz NOT NULL DEFAULT now(),++ follow_up_of text COLLATE "C" -- the attempt whose end started this one++);++CREATE UNIQUE INDEX IF NOT EXISTS codes_attempts_task_n ON codes_attempts (task_id, n);++-- One ending, one follow-up: the unique index is the fence (the second insert is a no-op).++CREATE UNIQUE INDEX IF NOT EXISTS codes_attempts_follow_up ON codes_attempts (follow_up_of) WHERE follow_up_of IS NOT NULL;++-- A Run job is one attempt: the job of an attempt is found by it.++CREATE INDEX IF NOT EXISTS codes_attempts_job ON codes_attempts (job_id) WHERE job_id IS NOT NULL;++CREATE INDEX IF NOT EXISTS codes_attempts_task ON codes_attempts (task_id, n);++++ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS outcome text COLLATE "C"; -- done | unfinished | stopped++ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS outcome_note text COLLATE "C"; -- what is left, in plain words++ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS attempts_count integer NOT NULL DEFAULT 0;++ALTER TABLE codes_session_lines ADD COLUMN IF NOT EXISTS attempt_n integer;++++-- Every task that exists becomes attempt 1 of itself. The id is the task's own id hashed into the ULID alphabet++-- (26 characters of Crockford base32), so the same task always gets the same attempt and a rerun of the migration++-- changes nothing.++INSERT INTO codes_attempts (id, task_id, n, agent, job_id, state, reason, why, kind, started_at, finished_at, created_at)++SELECT 'att_' || upper(substr(md5(t.id), 1, 26)),++ t.id,++ 1,++ t.agent,++ t.job_id,++ CASE t.state++ WHEN 'queued' THEN 'queued'++ WHEN 'placed' THEN 'queued'++ WHEN 'running' THEN 'running'++ WHEN 'succeeded' THEN 'done'++ WHEN 'cancelled' THEN 'stopped'++ ELSE 'failed'++ END,++ CASE t.state++ WHEN 'cancelled' THEN 'stopped'++ WHEN 'expired' THEN 'capacity'++ WHEN 'failed' THEN 'error'++ ELSE NULL++ END,++ CASE t.state++ WHEN 'succeeded' THEN 'the first attempt of this task finished the work'++ WHEN 'cancelled' THEN 'the first attempt of this task was stopped on purpose'++ WHEN 'expired' THEN 'no machine took the first attempt in time'++ WHEN 'failed' THEN 'the first attempt of this task failed'++ WHEN 'running' THEN 'the first attempt of this task is running'++ ELSE 'the first attempt of this task is waiting for a machine'++ END,++ 'first',++ t.started_at,++ t.finished_at,++ to_timestamp(GREATEST(t.created_at, 0) / 1000.0)++FROM run_tasks t++WHERE NOT EXISTS (SELECT 1 FROM codes_attempts a WHERE a.task_id = t.id);++++UPDATE run_tasks t++SET outcome = CASE t.state++ WHEN 'succeeded' THEN 'done'++ WHEN 'cancelled' THEN 'stopped'++ WHEN 'failed' THEN 'unfinished'++ WHEN 'expired' THEN 'unfinished'++ ELSE NULL++ END,++ outcome_note = CASE t.state++ WHEN 'succeeded' THEN 'the task finished before attempts were counted'++ WHEN 'cancelled' THEN 'the task was stopped on purpose'++ WHEN 'failed' THEN 'the attempt failed; what it left is in the session'++ WHEN 'expired' THEN 'no machine took the task in time'++ ELSE NULL++ END,++ attempts_count = 1++WHERE t.attempts_count = 0;++++UPDATE codes_session_lines SET attempt_n = 1 WHERE attempt_n IS NULL;@@@ -15,7 +15,7 @@"express-rate-limit": "^7.5.0","helmet": "^8.1.0","iovalkey": "^0.4.0",-- "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",++ "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0","openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0","openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0","pg": "^8.23.0"@@@ -1149,9 +1149,9 @@}},"node_modules/openvibe-contracts": {-- "version": "0.232.0",-- "resolved": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",-- "integrity": "sha512-6HI5nn//nAUEDdG4xRTAlcb40oc/6yO1nx7R5exGZ6Bo3mL5fJ5szw9IONddoLN6qJsO1CXzB5TzTqkBAh3OzQ==",++ "version": "0.243.0",++ "resolved": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0",++ "integrity": "sha512-ZFrwTpgALUEOIvdYi2r654murdEwludenAZ5u3CuvPMVcYuu/3kC+J20jVeqPaCFKYkzRzLyO8fH/iOdh44aYg==","license": "AGPL-3.0-only","dependencies": {"ajv": "^8.17.1",@@@ -30,7 +30,7 @@"express-rate-limit": "^7.5.0","helmet": "^8.1.0","iovalkey": "^0.4.0",-- "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",++ "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0","openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0","openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0","pg": "^8.23.0"@@@ -32,6 +32,7 @@ const { createTasksApi } = require('./http/tasks-api');const { createActionsApi } = require('./http/actions-api');const { createRunClient } = require('./domain/run-client');const { createRunTasks } = require('./domain/run-tasks');++const { createAttempts } = require('./domain/attempts');const { createSessionLines } = require('./domain/session-lines');const { createOutbox, createPublisher } = require('./events/outbox');const { createEventsConsumer } = require('./events-consumer');@@@ -66,8 +67,10 @@ async function createApp(opts = {}) {const sessions = createSessionLines({ store, config, log });const outbox = createOutbox({ store, log });const run = opts.run || createRunClient({ config, fetchImpl: fetchImpl || globalThis.fetch });-- const tasks = createRunTasks({ store, run, config, log, sessions, outbox });-- const ctx = { config, store, keys, sso, harnesses, log, sessions, outbox, run, tasks };++ // The task's attempts (stage C2): every Run job is one attempt, and the engine decides what happens next.++ const attempts = createAttempts({ store, config, log, harnesses });++ const tasks = createRunTasks({ store, run, config, log, sessions, outbox, attempts });++ const ctx = { config, store, keys, sso, harnesses, log, sessions, outbox, run, tasks, attempts };ctx.forgejo = createForgejoClient({ config, log });const app = express();@@@ -81,6 +81,18 @@ function load(env = process.env) {// OpenVibe.Run (plan T16): a coding task on the person's own machine is a Run job Codes starts for them.run: { url: trim(env.CODES_RUN_URL || 'http://127.0.0.1:4920') },++ // A task's own budget (engine/budget.js): steps (tool calls), minutes (wall clock) and dollars. A limit that is++ // not set is no limit — 0 means "no limit", never a zero budget. Steps and minutes buy one wrap-up turn.++ budget: {++ steps: int(env.CODES_TASK_STEPS, 0),++ minutes: int(env.CODES_TASK_MINUTES, 0),++ usd: Number(env.CODES_TASK_USD || 0) || 0,++ },++ // The task's timeline (stage C2): how many automatic continues one task may have before it needs a person.++ task: { maxContinues: Math.max(0, int(env.CODES_TASK_MAX_CONTINUES, 2)) },++ // The event stream of a task (GET /api/v1/tasks/:id/events): the heartbeat that keeps a connection alive.++ sse: { heartbeatMs: Math.max(1000, int(env.CODES_SSE_HEARTBEAT_MS, 15_000)) },++};}nnew file mode 100644@@@ -0,0 +1,418 @@++'use strict';++++/**++ * A coding task's attempts (codes_attempts): the timeline of one task.++ *++ * A continue, a switch to another agent and a wrap-up are attempts on one task's timeline, never new tasks with++ * cryptic names (the portal's "-c1"/"-r2" confusion of 10-30). Every Run job is one attempt: the attempt carries the++ * job, the agent, the machine and the harness session a resume needs.++ *++ * The engine decides what happens next (engine/chain.js outcome/nextStep, engine/budget.js wrapUp): this module turns++ * its decision into a plan — a follow-up attempt, or the task's outcome in plain words. It never starts a Run job++ * itself (server/domain/run-tasks.js does that, outside a transaction) and it never decides twice for one ending:++ * the follow-up's follow_up_of is unique, so one ending can start only one follow-up.++ *++ * createAttempts({ store, config, log, harnesses })++ * list(taskId) the attempts of a task, in order++ * get(id) one attempt++ * create({ task, …, followUpOf }) one attempt; the same followUpOf answers the same attempt (the fence)++ * bind(id, { job_id, state }) the Run job of an attempt++ * end(id, { state, reason, … }) an attempt ended: its state, its reason, its cost++ * decide({ task, attempts, ending, result }) the engine's decision, as a plan++ * present(attempt) the attempt as the API shows it++ */++++const { chain, budget, router } = require('../../engine');++++/** Codes' harness ids and the engine's short runner ids for the same harness. */++const RUNNER_OF = { 'claude-code': 'claude', codex: 'cx', 'command-code': 'cc', opencode: 'oc', deepseek: 'dsh' };++const HARNESS_OF = { claude: 'claude-code', cx: 'codex', cc: 'command-code', oc: 'opencode', dsh: 'deepseek' };++++/** The order a runner switch tries the harnesses in (the ladder's order, flattened). */++const FALLBACK_ORDER = ['cx', 'claude', 'cc', 'oc', 'dsh'];++++/** Which harnesses may take which kind of task, cheapest first (the engine's ladder, Codes' harnesses). */++const TIER_ORDER = {++ basic: ['opencode', 'command-code', 'deepseek'],++ research: ['codex', 'opencode', 'deepseek'],++ standard: ['codex', 'claude-code', 'command-code', 'opencode', 'deepseek'],++ hard: ['codex', 'claude-code', 'deepseek'],++ design: ['codex', 'claude-code', 'opencode'],++};++const TIER_EFFORT = { basic: 'low', research: 'medium', standard: 'high', hard: 'high', design: 'high' };++++/** The attempt's own states, and the reasons it may have ended. */++const STATES = ['queued', 'running', 'done', 'failed', 'stopped', 'budget-stopped', 'interrupted'];++const REASONS = ['limit', 'auth', 'capacity', 'refused', 'error', 'budget', 'stopped', 'hung', 'context'];++const KINDS = ['first', 'continue', 'switch', 'wrap-up', 'restart', 'manual'];++++/** The engine's own words for a reason: the timeline reads like a person wrote it. */++const REASON_TEXT = chain.REASON_TEXT;++++/**++ * The text the engine classifies a reason from. The engine reads the reason of a failed attempt out of what it said++ * (engine/chain.js reason), so the text has to say the same thing our reason column says.++ */++const ERROR_TEXT = {++ limit: 'usage limit reached',++ auth: 'authentication failed: log in again',++ capacity: 'the model was at capacity',++ refused: 'turned off in the agent portal',++ error: 'the work failed',++ budget: 'stopped at its limit',++ stopped: 'stopped on purpose',++ context: 'the context window is too big',++ hung: 'it went quiet for too long',++};++++/** The engine's status for our state (the two vocabularies are the same, except 'queued' and 'running'). */++const ENGINE_STATE = {++ queued: 'queued', running: 'running', done: 'done', failed: 'failed',++ stopped: 'stopped', 'budget-stopped': 'budget-stopped', interrupted: 'interrupted',++};++++/** The attempt's state for a task's state (a task from before the attempts still has one). */++const TASK_STATE = {++ queued: 'queued', placed: 'queued', running: 'running', succeeded: 'done',++ failed: 'failed', cancelled: 'stopped', expired: 'failed',++};++++const iso = (v) => {++ if (!v) return null;++ const t = v instanceof Date ? v.getTime() : Date.parse(String(v));++ return Number.isFinite(t) ? new Date(t).toISOString() : null;++};++const clip = (s, n) => String(s == null ? '' : s).slice(0, n);++const runnerOf = (agent) => RUNNER_OF[agent] || (agent || 'auto');++const wordsOf = (reason) => REASON_TEXT[reason] || REASON_TEXT.error;++++function createAttempts({ store, config = {}, log = console, harnesses = null }) {++ const db = store.db;++ const catalog = harnesses || require('./harnesses').createHarnesses();++ // How many automatic continues one task may have before it needs a person (CODES_TASK_MAX_CONTINUES, two).++ const wanted = config.task ? config.task.maxContinues : undefined;++ const maxContinues = Number.isFinite(Number(wanted)) ? Math.max(0, Number(wanted)) : 2;++++ const nameOf = (agent) => {++ const offer = agent ? catalog.get(agent) : null;++ return offer ? offer.name : (agent || 'the agent');++ };++++ /** The attempts of a task, in order. A task written before the attempts (or by another path) still has a first one. */++ async function list(taskId) {++ if (!taskId) return [];++ const rows = await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);++ if (rows.length) return rows;++ const task = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [taskId]);++ if (!task) return [];++ await db.exec(++ `INSERT INTO codes_attempts (id, task_id, n, agent, job_id, state, why, kind, started_at, finished_at, created_at)++ VALUES ($1, $2, 1, $3, $4, $5, $6, 'first', $7::timestamptz, $8::timestamptz, $9::timestamptz)++ ON CONFLICT DO NOTHING`,++ [store.newId('att'), task.id, task.agent || null, task.job_id || null, TASK_STATE[task.state] || 'failed',++ 'the first attempt of this task', task.started_at || null, task.finished_at || null, store.iso()],++ );++ return await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);++ }++++ /** One attempt. */++ async function get(id) {++ if (!id) return null;++ return await db.maybe('SELECT * FROM codes_attempts WHERE id = $1', [id]);++ }++++ /** The attempt a job belongs to, or null. */++ async function byJob(jobId) {++ if (!jobId) return null;++ return await db.maybe('SELECT * FROM codes_attempts WHERE job_id = $1', [jobId]);++ }++++ /**++ * One attempt. The fence: follow_up_of is unique, so an ending that starts a follow-up twice answers the same++ * attempt the second time and starts no second job. created tells the caller which of the two it was.++ */++ async function create({ task, agent = null, model = null, machine = null, session = null, why = null, kind = 'first', followUpOf = null, state = 'queued' }) {++ if (!task || !task.id) throw new Error('an attempt belongs to a task');++ for (let tries = 0; tries < 5; tries++) {++ const n = await db.one('SELECT COALESCE(MAX(n), 0) + 1 AS n FROM codes_attempts WHERE task_id = $1', [task.id]);++ const id = store.newId('att');++ const inserted = await db.maybe(++ `INSERT INTO codes_attempts (id, task_id, n, agent, model, machine, session, state, why, kind, follow_up_of, created_at)++ VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12::timestamptz)++ ON CONFLICT DO NOTHING RETURNING *`,++ [id, task.id, n.n, agent, model, machine, session, state, why, kind, followUpOf, store.iso()],++ );++ if (inserted) {++ await db.exec('UPDATE run_tasks SET attempts_count = $2 WHERE id = $1', [task.id, n.n]);++ return { attempt: inserted, created: true };++ }++ // Someone else wrote it: the same follow-up answers the same attempt.++ const seen = followUpOf ? await db.maybe('SELECT * FROM codes_attempts WHERE follow_up_of = $1', [followUpOf]) : null;++ if (seen) return { attempt: seen, created: false };++ }++ throw new Error('the attempt could not be written');++ }++++ /** The Run job of an attempt. */++ async function bind(id, { job_id = null, state = null, started_at = null, machine = null, model = null } = {}) {++ return await db.one(++ `UPDATE codes_attempts SET job_id = COALESCE($2, job_id),++ state = COALESCE($3, state),++ machine = COALESCE($4, machine),++ model = COALESCE($5, model),++ started_at = COALESCE($6::timestamptz, started_at)++ WHERE id = $1 RETURNING *`,++ [id, job_id, state, machine, model, started_at],++ );++ }++++ /** An attempt ended: its state, why, what it cost. */++ async function end(id, { state = 'failed', reason = null, why = null, cost_usd = null, tokens = null, steps = null, session = null, model = null, machine = null, finished_at = null } = {}) {++ return await db.one(++ `UPDATE codes_attempts SET state = $2,++ reason = COALESCE($3, reason),++ why = COALESCE($4, why),++ cost_usd = COALESCE($5::numeric, cost_usd),++ tokens = COALESCE($6, tokens),++ steps = COALESCE($7, steps),++ session = COALESCE($8, session),++ model = COALESCE($9, model),++ machine = COALESCE($10, machine),++ finished_at = COALESCE($11::timestamptz, finished_at)++ WHERE id = $1 RETURNING *`,++ [id, state, reason, why, cost_usd == null ? null : String(cost_usd), tokens, steps, session, model, machine, finished_at || store.iso()],++ );++ }++++ /** The engine's own vocabulary for Codes' harnesses: the ladder, the health and the catalogs it routes with. */++ function routing() {++ const subagents = {}; const models = {}; const runners = [];++ for (const [harnessId, runner] of Object.entries(RUNNER_OF)) {++ const offer = catalog.get(harnessId);++ if (!offer) continue;++ subagents[runner] = { enabled: true, max: 4, model: null };++ models[runner] = (offer.agents || []).map((a) => a.model).filter(Boolean);++ runners.push(runner);++ }++ const health = {};++ for (const r of runners) health[r] = { state: 'ok' };++ const ladder = {};++ for (const tier of router.TIERS) {++ ladder[tier] = (TIER_ORDER[tier] || []).map((id) => {++ const runner = RUNNER_OF[id];++ const offer = catalog.get(id);++ const agent = offer && offer.agents && offer.agents[0];++ return [runner, agent ? agent.model : null, TIER_EFFORT[tier]];++ }).filter(([runner, model]) => subagents[runner] && (!model || models[runner].includes(model)));++ }++ return {++ cfg: { subagents, budget: config.budget || {}, fallback: { order: FALLBACK_ORDER.filter((r) => subagents[r]) } },++ health, counts: {}, models, runners, ladder,++ };++ }++++ /**++ * The next agent for a task, from the engine's router (engine/router.js): the tier of the brief decides the++ * ladder, and the first harness of it that may take the task takes it. The agent that just stopped is skipped.++ */++ function nextAgent(prompt, ctx, exclude) {++ const usable = (runner) => {++ const sub = ctx.cfg.subagents[runner];++ if (!sub || !sub.enabled) return false;++ if ((ctx.counts[runner] || 0) >= sub.max) return false;++ const h = ctx.health[runner];++ return !!h && h.state === 'ok';++ };++ const picked = router.pick(prompt, ctx);++ const tier = picked ? picked.tier : router.classify(prompt).tier;++ const ladder = ctx.ladder[tier] || [];++ const shape = (runner, model, effort) => ({++ runner,++ agent: HARNESS_OF[runner] || null,++ model: model || null,++ effort: router.fitEffort(runner, effort),++ tier,++ why: picked ? picked.why : `${tier} → ${runner}`,++ });++ if (picked && picked.runner !== exclude && usable(picked.runner)) return shape(picked.runner, picked.model, picked.effort);++ for (const [runner, model, effort] of ladder) {++ if (runner === exclude || !usable(runner)) continue;++ const m = model && ctx.models[runner] && !ctx.models[runner].includes(model) ? null : model;++ return shape(runner, m, effort);++ }++ return null;++ }++++ /** One attempt as the engine sees it. */++ function engineAttempt(a, all, facts = null) {++ const n = Number(a.n) || 1;++ const reason = a.reason && REASONS.includes(a.reason) ? a.reason : null;++ const next = all.find((x) => x.follow_up_of && x.follow_up_of === a.id);++ const prev = all.find((x) => Number(x.n) === n - 1);++ const created = a.created_at ? Date.parse(String(a.created_at)) : 0;++ return {++ key: a.id,++ runner: runnerOf(a.agent),++ name: n === 1 ? 'a' : `a-r${n}`,++ runnerName: nameOf(a.agent),++ status: ENGINE_STATE[a.state] || 'failed',++ // The order the engine reads the chain in: the attempt's place, never a tie.++ updated: created + n,++ attempt: n,++ error: reason ? ERROR_TEXT[reason] : null,++ // The engine reads the reason of a failed attempt out of what it said; Codes' attempts carry their own++ // reason, and the engine only reads the runners it knows, so it is passed on.++ reason: reason || null,++ next: next ? next.id : null,++ prev: prev ? prev.id : null,++ parent: a.follow_up_of || null,++ stopped: a.state === 'stopped',++ budget: reason === 'budget' ? { hit: 'steps' } : undefined,++ wrapUp: a.kind === 'wrap-up',++ hung: reason === 'hung',++ facts: {++ session: a.session || null,++ apiError: (facts && facts.apiError) || (reason ? ERROR_TEXT[reason] : null),++ hasFinal: !!(facts && facts.hasFinal),++ finalText: (facts && facts.finalText) || '',++ },++ };++ }++++ /**++ * What the job's log says about the attempt: its last words (the engine reads them to know whether the agent++ * said it was done) and the error it ended with.++ */++ function factsOf(result) {++ const d = result && typeof result === 'object' && !Array.isArray(result) ? result : {};++ const log = Array.isArray(d.log) ? d.log : [];++ let hasFinal = false; let finalText = ''; let apiError = '';++ for (const line of log) {++ if (!line || typeof line !== 'object') continue;++ if (line.t === 'done') {++ hasFinal = true;++ if (line.text) finalText = String(line.text);++ if (line.ok === false && !apiError) apiError = finalText || 'the agent ended without an answer';++ } else if (line.t === 'note' && line.level === 'error' && !apiError) {++ apiError = String(line.text || '');++ }++ }++ return { hasFinal, finalText, apiError };++ }++++ /** The note a continued attempt gets: it keeps the work, it does not redo it. */++ function continuationNote(ending, words) {++ return `\n\n---\n## Continuation (written by openvibe.codes)\n\nAn earlier attempt${ending.agent ? ` on ${nameOf(ending.agent)}` : ''} ${words}. Keep the work it finished, do not redo it, and continue from where it stopped. If the task is already complete, check it and report that.\n`;++ }++++ /** The note a new agent gets: what changed, what it said, where it stopped. */++ function handoverNote(ending, result, words) {++ const d = result && typeof result === 'object' ? result : {};++ const files = Array.isArray(d.files) ? d.files.map((f) => f && f.path).filter(Boolean).slice(0, 20) : [];++ const summary = clip(d.summary || '', 1200);++ return `\n\n---\n## Hand-over (written by openvibe.codes)\n\n${nameOf(ending.agent)} ${words}. Before it stopped it:\n`++ + `- changed: ${files.length ? files.join(', ') : 'nothing yet'}\n`++ + `- said: ${summary || 'nothing'}\n`++ + `- stopped at: ${words}\n\n`++ + 'Keep the work it did, do not redo it, and finish the task.\n';++ }++++ /**++ * What happens after an attempt ends: the engine decides, this turns the decision into a plan.++ *++ * @param {object} what++ * @param {object} what.task the task row++ * @param {object[]} what.attempts every attempt of the task, in order++ * @param {object} what.ending the attempt that ended++ * @param {object} [what.result] the job's result (what it changed and what it said)++ * @returns {object} the plan: an outcome ('done', 'unfinished', 'stopped') or a follow-up attempt++ */++ function decide({ task, attempts, ending, result = null }) {++ const ctx = routing();++ const upto = (attempts || []).filter((a) => Number(a.n) <= Number(ending.n));++ const facts = factsOf(result);++ const engineAttempts = upto.map((a) => engineAttempt(a, upto, a.id === ending.id ? facts : null));++ const out = chain.outcome(engineAttempts, ctx);++ const step = out.step || { action: 'none' };++ const reason = ending.reason || out.reason || 'error';++ const words = wordsOf(reason);++ const continues = upto.filter((a) => a.kind === 'continue').length;++++ if (step.action === 'done') return { outcome: 'done', note: 'the task finished' };++ if (step.action === 'stop') return { outcome: 'stopped', note: 'the task was stopped on purpose' };++ if (step.action === 'wait') return { outcome: null, note: null };++++ if (step.action === 'wrap-up') {++ const hit = step.hit || 'steps';++ return {++ kind: 'wrap-up', agent: ending.agent, model: ending.model, session: ending.session || null,++ prompt: String(task.prompt || '') + budget.wrapUpPrompt(hit),++ why: `one last turn to land the work after the ${hit} limit`,++ };++ }++++ if (step.action === 'continue') {++ if (continues >= maxContinues) {++ return { outcome: 'unfinished', note: `the task was carried on ${continues} times already; it needs a person` };++ }++ return {++ kind: 'continue', agent: ending.agent, model: ending.model, session: step.session || ending.session || null,++ prompt: String(task.prompt || '') + continuationNote(ending, words),++ why: `carried on in the same conversation: ${words}`,++ };++ }++++ if (step.action === 'restart') {++ return {++ kind: 'restart', agent: ending.agent, model: ending.model, session: null,++ prompt: String(task.prompt || '') + continuationNote(ending, words),++ why: `started again with a fresh session: ${words}`,++ };++ }++++ if (step.action === 'switch') {++ const choice = nextAgent(task.prompt, ctx, runnerOf(ending.agent));++ const agent = choice ? choice.agent : (step.runner ? HARNESS_OF[step.runner] : null);++ if (!agent) return { outcome: 'unfinished', note: `no other agent could take it: ${step.why || words}` };++ return {++ kind: 'switch', agent, model: choice ? choice.model : (step.model || null), session: null,++ prompt: String(task.prompt || '') + handoverNote(ending, result, words),++ why: `${nameOf(agent)} took over: ${words}`,++ };++ }++++ return { outcome: 'unfinished', note: `${step.why || 'the task stopped'}. What is left is in the session.` };++ }++++ /** An attempt as the API and the page show it. */++ function present(a) {++ if (!a) return null;++ return {++ id: a.id,++ n: Number(a.n) || 0,++ agent: a.agent || null,++ model: a.model || null,++ machine: a.machine || null,++ job_id: a.job_id || null,++ session: a.session || null,++ state: a.state,++ reason: a.reason || null,++ why: a.why || null,++ kind: a.kind || 'first',++ cost_usd: a.cost_usd == null ? null : Number(a.cost_usd),++ tokens: a.tokens == null ? null : Number(a.tokens),++ steps: a.steps == null ? null : Number(a.steps),++ started_at: iso(a.started_at),++ finished_at: iso(a.finished_at),++ created_at: iso(a.created_at),++ end_why: a.reason ? wordsOf(a.reason) : null,++ };++ }++++ return {++ STATES, REASONS, KINDS, REASON_TEXT, RUNNER_OF, HARNESS_OF,++ list, get, byJob, create, bind, end, decide, present, routing, nextAgent, nameOf, wordsOf,++ };++}++++module.exports = { createAttempts, RUNNER_OF, HARNESS_OF, STATES, REASONS, KINDS, REASON_TEXT, ERROR_TEXT };@@@ -4,10 +4,13 @@* Coding tasks on a person's own machine (run_tasks, plan T16): what the /run pages (server/http/run.js) and the* service API (server/http/tasks-api.js, codes.task.*) share.*-- * createRunTasks({ store, run, config, log, sessions, outbox })++ * createRunTasks({ store, run, config, log, sessions, outbox, attempts })* start({ subject, project, args, via, key }) → the row: queued (Run took the job) or failed (Run's reason)* refresh(row, { force }) → the row, read from Run only as a fallback after 10 quiet minutes* applyEvent(row, event) → a run.job.* delivery: state, timings, result, event, in one tx++ * afterEnd(taskId) → the engine's decision after an attempt ended++ * continueTask(row, { note, agent }) → the next attempt now (a switch when an agent is named)++ * stopTask(row) → stop the running attempt; nothing carries it on* latestProject(subject) → the project of the person's latest task, or null* present(row) → codes.task@1* TERMINAL, STATE_LABEL, STALE_MS@@@ -15,8 +18,15 @@* Every state change writes its codes.task.* outbox row in the same transaction; a terminal state also back-fills* the session's missing lines and marks it complete. A task's result comes from the person's machine: present() passes* on only the contract's fields, each checked and cut to its limit.++ *++ * Stage C2: every Run job is one attempt on the task's timeline (server/domain/attempts.js). When an attempt ends,++ * the engine decides what happens next — done, a continue in the same session, a switch to another agent, a wrap-up or++ * an unfinished task — and this module carries the decision out. One ending starts at most one follow-up: the++ * follow-up's follow_up_of is unique, so two racing ticks never carry one job on twice.*/const { RunError } = require('./run-client');++const { chain, budget } = require('../../engine');++const { parseArgs } = require('../../harness/job');const TERMINAL = new Set(['succeeded', 'failed', 'cancelled', 'expired']);const STATE_LABEL = {@@@ -85,12 +95,60 @@ function errorOf(job) {return `${job.error.code || 'run.error'}${job.error.detail ? `: ${job.error.detail}` : ''}`.slice(0, 1000);}--function createRunTasks({ store, run, config, log = console, sessions = null, outbox = null }) {++/** The machine a job ran on: the worker's device id, or the hosted sandbox. */++function machineOf(job) {++ if (!job || typeof job !== 'object') return null;++ const node = job.placement && typeof job.placement === 'object' ? job.placement.node : null;++ if (typeof node === 'string' && node) return node.slice(0, 128);++ return typeof job.machine === 'string' && job.machine ? job.machine.slice(0, 128) : null;++}++++/** The attempt's own state for a Run job's state. */++const ATTEMPT_STATE = { queued: 'queued', placed: 'queued', running: 'running', succeeded: 'done', failed: 'failed', cancelled: 'stopped', expired: 'failed' };++++/** The reason an attempt ended: the engine's own words (engine/chain.js reason), plus the two it reads from the text. */++function reasonOf(state, text) {++ if (state === 'cancelled') return 'stopped';++ if (state === 'expired') return 'capacity';++ const s = String(text || '');++ if (/went quiet|hung|stale/i.test(s)) return 'hung';++ if (/prompt is too long|context length|maximum context|context window|too many tokens|input too long/i.test(s)) return 'context';++ const why = chain.reason(s);++ return why === 'unknown' ? 'error' : why;++}++++/** What an attempt spent: its cost, its tokens, its steps and the session a resume needs (from the job's result log). */++function usageOf(result) {++ const d = asObject(result) || {};++ const log = Array.isArray(d.log) ? d.log : [];++ let tokens = 0; let steps = 0; let session = null; let model = null;++ for (const line of log) {++ if (!line || typeof line !== 'object') continue;++ if (line.t === 'usage') tokens += (Number(line.input_tokens) || 0) + (Number(line.output_tokens) || 0);++ else if (line.t === 'tool') steps++;++ else if (line.t === 'agent') {++ if (!session && line.session) session = String(line.session);++ if (!model && line.model) model = String(line.model);++ }++ }++ return {++ cost_usd: typeof d.cost_usd === 'number' && Number.isFinite(d.cost_usd) ? d.cost_usd : null,++ tokens: tokens || null,++ steps: steps || null,++ session: session || null,++ model: model || null,++ };++}++++function createRunTasks({ store, run, config, log = console, sessions = null, outbox = null, attempts = null }) {const db = store.db;const who = (row) => ({ subject: row.subject, project: row.project_id });// Lazily, so a caller that only reads tasks does not need the session/outbox modules (and there is no cycle).const sessionLines = sessions || require('./session-lines').createSessionLines({ store, config, log });const taskOutbox = outbox || require('../events/outbox').createOutbox({ store, log });++ const attemptStore = attempts || require('./attempts').createAttempts({ store, config, log });++ /** An attempt's state when it ended: a limit reached is a budget-stop, never a plain failure. */++ const ENDED = new Set(['done', 'failed', 'stopped', 'budget-stopped', 'interrupted']);/*** One state change: update the row, write its codes.task.* outbox row in the same transaction, and on a terminal@@@ -121,36 +179,79 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou[current.id, state, result == null ? null : JSON.stringify(result), error, startedAt, finishedAt, store.now()],);if (changed) await taskOutbox.emitTask(updated, current.state);++ // Every Run job is one attempt: the attempt of this job follows the job.++ const attempt = await attemptStore.byJob(current.job_id);if (TERMINAL.has(updated.state) && !updated.session_complete_at) {-- await sessionLines.backfillAndComplete(updated, updated.result);++ await sessionLines.backfillAndComplete(updated, updated.result, attempt ? Number(attempt.n) : null);++ }++ if (attempt) {++ if (TERMINAL.has(state)) {++ const usage = usageOf(updated.result);++ const why = reasonOf(state, updated.error);++ const hit = budgetHit(attempt, usage, why);++ await attemptStore.end(attempt.id, {++ state: hit ? 'budget-stopped' : (ATTEMPT_STATE[state] || 'failed'),++ reason: hit ? 'budget' : why,++ cost_usd: usage.cost_usd, tokens: usage.tokens, steps: usage.steps, session: usage.session,++ model: usage.model, machine: machineOf(next.job),++ finished_at: updated.finished_at || store.iso(),++ });++ } else {++ await attemptStore.bind(attempt.id, {++ state: ATTEMPT_STATE[state] || 'queued', started_at: updated.started_at || null, machine: machineOf(next.job),++ });++ }}return updated;});}++ /**++ * The limit an attempt reached, if any (engine/budget.js): steps, minutes or dollars. A provider's own limit++ * (usage limit, signed out) is not the owner's budget and never buys a wrap-up turn.++ */++ function budgetHit(attempt, usage, reason) {++ if (!['error', 'hung', 'context'].includes(reason)) return null;++ const limits = budget.limitsFor(config, attempt.agent, null);++ if (!limits.steps && !limits.minutes && !limits.usd) return null;++ const started = attempt.started_at ? Date.parse(String(attempt.started_at)) : 0;++ const finished = attempt.finished_at ? Date.parse(String(attempt.finished_at)) : 0;++ const minutes = finished > started ? (finished - started) / 60000 : null;++ return budget.evaluate({ steps: usage.steps || 0, minutes, usd: usage.cost_usd }, limits).hit;++ }++async function start({ subject, project, args, via = null, key = null, actorRef = null }) {const id = store.newId('tsk');const at = store.now();const actorRefValue = actorRef && typeof actorRef === 'object' && !Array.isArray(actorRef) ? JSON.stringify(actorRef) : null;++ let inserted;await store.tx(async () => {-- const inserted = await db.one(`INSERT INTO run_tasks (id, subject, project_id, repo, ref, prompt, agent, push, pull_request, state, via, idempotency_key, actor_ref, created_at, updated_at)++ inserted = await db.one(`INSERT INTO run_tasks (id, subject, project_id, repo, ref, prompt, agent, push, pull_request, state, via, idempotency_key, actor_ref, created_at, updated_at)VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, 'queued', $10, $11, $12::jsonb, $13, $13) RETURNING *`,[id, subject, project, args.repo, args.ref, args.prompt, args.agent, !!args.push, !!(args.push && args.push.pull_request), via, key, actorRefValue, at]);await taskOutbox.emitTask(inserted, null);return inserted;});++ // The task's first attempt: every Run job is one attempt, and this one is the task as it was asked.++ const first = await attemptStore.create({++ task: inserted, agent: args.agent || null, model: args.model || null,++ why: args.agent ? `the first try, with ${attemptStore.nameOf(args.agent)}` : 'the first try, with the agent that fits the task',++ kind: 'first', state: 'queued',++ });let job;try {job = await run.submit({ subject, project, args, key: id });} catch (e) {const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;++ await attemptStore.end(first.attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });return await store.tx(async () => {-- const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, finished_at = $3::timestamptz, updated_at = $4-- WHERE id = $1 RETURNING *`, [id, why, store.iso(), store.now()]);++ const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, outcome = 'unfinished', outcome_note = $3, finished_at = $4::timestamptz, updated_at = $5++ WHERE id = $1 RETURNING *`, [id, why, `the task could not start: ${why}`, store.iso(), store.now()]);await taskOutbox.emitTask(failed, 'queued');return failed;});}++ await attemptStore.bind(first.attempt.id, { job_id: job.id, state: 'queued' });return await store.tx(async () => {const state = STATES.has(job.state) ? job.state : 'queued';const startedAt = job.timings && job.timings.started_at ? job.timings.started_at : null;@@@ -165,7 +266,7 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ouasync function applyJob(row, job) {if (!job || !STATES.has(job.state)) return row;const result = TERMINAL.has(job.state) ? (asObject(job.result) || null) : undefined;-- return await recordState(row, {++ const updated = await recordState(row, {state: job.state,result,error: TERMINAL.has(job.state) ? errorOf(job) : undefined,@@@ -173,6 +274,8 @@ function createRunTasks({ store, run, config, log = console, sessions = null, oufinishedAt: job.timings && job.timings.finished_at,job,});++ // The attempt ended: the engine decides, and the row carries the outcome (or the next attempt).++ return (await afterEnd(updated.id)) || updated;}/**@@@ -180,7 +283,15 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou* for a missing delivery, not page-view polling). force reads now, for a cancel or an explicit check.*/async function refresh(row, { force = false } = {}) {-- if (!row.job_id || TERMINAL.has(row.state)) return row;++ if (!row.job_id) return row;++ if (TERMINAL.has(row.state)) {++ // The task has ended. Its decision (the outcome, or the attempt that follows) is made once, when the++ // attempt ended; if that failed, it is made here, as the fallback reads Run after 10 quiet minutes.++ if (!row.outcome && (force || store.now() - Number(row.updated_at) >= STALE_MS)) {++ return (await afterEnd(row.id)) || row;++ }++ return row;++ }if (!force && store.now() - Number(row.updated_at) < STALE_MS) return row;let job;try { job = await run.read(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] task ${row.id}: ${e.message}`); return row; }@@@ -201,6 +312,160 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou});}++ /**++ * An attempt ended: the engine decides what happens next (server/domain/attempts.js) and this carries the++ * decision out — a follow-up attempt, or the task's outcome in plain words.++ *++ * Calling it twice for one ending changes nothing: the follow-up's follow_up_of is unique, so the second call++ * finds the attempt the first one wrote and starts no second job (the portal's 10-30 bug, where two racing ticks++ * carried one job on twice).++ */++ async function afterEnd(taskId) {++ const row = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [taskId]);++ if (!row) return null;++ if (!TERMINAL.has(row.state)) return row;++ if (row.outcome === 'stopped') return row; // a person stopped it: nothing carries it on++ const attempts = await attemptStore.list(row.id);++ const ending = attempts[attempts.length - 1];++ if (!ending || !ENDED.has(ending.state)) return row;++ const plan = await attemptStore.decide({ task: row, attempts, ending, result: row.result });++ if (plan.outcome) {++ return await db.one('UPDATE run_tasks SET outcome = $2, outcome_note = $3, updated_at = $4 WHERE id = $1 RETURNING *',++ [row.id, plan.outcome, plan.note, store.now()]);++ }++ if (!plan.kind) return row;++ const done = await startFollowUp(row, plan, ending);++ return done.row || row;++ }++++ /**++ * The next attempt: a Run job for the task with the engine's plan in it. The attempt is written first and the job++ * second, so the fence (follow_up_of) is in place before anything is started.++ */++ async function startFollowUp(row, plan, ending) {++ const { attempt, created } = await attemptStore.create({++ task: row, agent: plan.agent || null, model: plan.model || null, session: plan.session || null,++ why: plan.why || null, kind: plan.kind || 'manual', followUpOf: ending.id, state: 'queued',++ });++ // Someone else already started the follow-up of this ending: no second job.++ if (!created) return { attempt, created: false, row };++ let args;++ try {++ args = followUpArgs(row, plan);++ } catch (e) {++ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });++ const failed = await db.one(`UPDATE run_tasks SET outcome = 'unfinished', outcome_note = $2, updated_at = $3 WHERE id = $1 RETURNING *`,++ [row.id, `the next attempt could not be prepared: ${e.message}`, store.now()]);++ return { attempt, created: true, row: failed, error: e.message };++ }++ let job;++ try {++ job = await run.submit({ subject: row.subject, project: row.project_id, args, key: attempt.id });++ } catch (e) {++ const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;++ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });++ const failed = await db.one(`UPDATE run_tasks SET outcome = 'unfinished', outcome_note = $2, updated_at = $3 WHERE id = $1 RETURNING *`,++ [row.id, `the next attempt could not start: ${why}`, store.now()]);++ return { attempt, created: true, row: failed, error: why };++ }++ await attemptStore.bind(attempt.id, { job_id: job.id, state: 'queued' });++ const updated = await db.one(++ `UPDATE run_tasks SET job_id = $2, state = 'queued', outcome = NULL, outcome_note = NULL, error = NULL,++ finished_at = NULL, updated_at = $3 WHERE id = $1 RETURNING *`,++ [row.id, job.id, store.now()]);++ await taskOutbox.emitTask(updated, row.state);++ return { attempt, created: true, job, row: updated };++ }++++ /**++ * The next attempt now: the owner carries the task on (with a note), or switches to another agent. The task's row++ * is locked while the attempt is written, so two presses start one attempt and not two.++ */++ async function continueTask(row, { note = null, agent = null } = {}) {++ if (!row || !row.id) return { error: 'no such task' };++ const attempts = await attemptStore.list(row.id);++ const ending = attempts[attempts.length - 1];++ if (!ending) return { error: 'the task has no attempt to carry on' };++ const text = note ? String(note).trim().slice(0, 4000) : '';++ const plan = agent++ ? {++ kind: 'switch', agent, model: null, session: null,++ prompt: `${row.prompt}\n\n---\n## Hand-over (written by openvibe.codes)\n\n${attemptStore.nameOf(agent)} takes over from ${attemptStore.nameOf(ending.agent)}, which ${attemptStore.wordsOf(ending.reason || 'error')}. Keep the work that was done, do not redo it, and finish the task.\n`,++ why: `${attemptStore.nameOf(agent)} took over at the owner's request`,++ }++ : {++ kind: 'manual', agent: ending.agent, model: ending.model, session: null,++ prompt: `${row.prompt}${text ? `\n\n---\n## A note from the owner\n\n${text}\n` : ''}`,++ why: text ? `carried on by the owner: ${text.slice(0, 200)}` : 'carried on by the owner',++ };++ let attempt = null;++ const locked = await store.tx(async () => {++ const current = await db.maybe('SELECT * FROM run_tasks WHERE id = $1 FOR UPDATE', [row.id]);++ if (!current) return { error: 'no such task' };++ if (!TERMINAL.has(current.state)) return { error: 'the task is still running' };++ if (current.outcome === 'stopped') return { error: 'the task was stopped on purpose' };++ const made = await attemptStore.create({++ task: current, agent: plan.agent, model: plan.model, session: null, why: plan.why, kind: plan.kind, state: 'queued',++ });++ attempt = made.attempt;++ await db.one(`UPDATE run_tasks SET state = 'queued', outcome = NULL, outcome_note = NULL, error = NULL, job_id = NULL, finished_at = NULL, updated_at = $2 WHERE id = $1 RETURNING *`,++ [row.id, store.now()]);++ return { row: current };++ });++ if (locked.error) return { error: locked.error };++ let job;++ try {++ job = await run.submit({ subject: row.subject, project: row.project_id, args: followUpArgs(row, plan), key: attempt.id });++ } catch (e) {++ const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;++ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });++ const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, outcome = 'unfinished', outcome_note = $3, finished_at = $4::timestamptz, updated_at = $5 WHERE id = $1 RETURNING *`,++ [row.id, why, `the next attempt could not start: ${why}`, store.iso(), store.now()]);++ return { error: why, row: failed, attempt };++ }++ await attemptStore.bind(attempt.id, { job_id: job.id, state: 'queued' });++ const updated = await db.one('UPDATE run_tasks SET job_id = $2, updated_at = $3 WHERE id = $1 RETURNING *', [row.id, job.id, store.now()]);++ await taskOutbox.emitTask(updated, row.state);++ return { attempt, job, row: updated };++ }++++ /** The job's arguments for a follow-up: the task's own repository, the plan's prompt and the session to resume. */++ function followUpArgs(row, plan) {++ return parseArgs(JSON.stringify({++ repo: row.repo, ref: row.ref || null, prompt: plan.prompt || row.prompt,++ agent: plan.agent || null, permission: 'edit',++ ...(row.push ? { push: { pull_request: !!row.pull_request } } : {}),++ ...(plan.session ? { resume: plan.session } : {}),++ }));++ }++++ /**++ * Stop the running attempt: the outcome is stopped and nothing carries it on. The task's row is locked, so a stop++ * and a follow-up never race.++ */++ async function stopTask(row) {++ if (!row || !row.id) return { error: 'no such task' };++ const attempts = await attemptStore.list(row.id);++ const running = attempts.slice().reverse().find((a) => !ENDED.has(a.state));++ if (row.job_id && !TERMINAL.has(row.state)) {++ try { await run.cancel(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] stop ${row.id}: ${e.message}`); }++ }++ const stopped = await store.tx(async () => {++ const current = await db.maybe('SELECT * FROM run_tasks WHERE id = $1 FOR UPDATE', [row.id]);++ if (!current) return { error: 'no such task' };++ if (TERMINAL.has(current.state) && current.outcome) return { error: 'the task has already ended', row: current };++ const updated = await db.one(++ `UPDATE run_tasks SET state = 'cancelled', outcome = 'stopped', outcome_note = $2,++ finished_at = COALESCE(finished_at, $3::timestamptz), updated_at = $4 WHERE id = $1 RETURNING *`,++ [row.id, 'the task was stopped on purpose', store.iso(), store.now()]);++ await taskOutbox.emitTask(updated, current.state);++ return { row: updated };++ });++ if (stopped.error) return stopped;++ if (running) await attemptStore.end(running.id, { state: 'stopped', reason: 'stopped', finished_at: store.iso() });++ return { row: stopped.row, attempt: running || null };++ }++async function latestProject(subject) {const row = await db.maybe('SELECT project_id FROM run_tasks WHERE subject = $1 ORDER BY created_at DESC LIMIT 1', [subject]);return row ? row.project_id : null;@@@ -215,7 +480,10 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou});}-- return { start, refresh, applyJob, applyEvent, recordState, latestProject, present, who };++ return {++ start, refresh, applyJob, applyEvent, recordState, afterEnd, continueTask, stopTask,++ latestProject, present, who, attempts: attemptStore,++ };}--module.exports = { createRunTasks, presentResult, TERMINAL, STATE_LABEL, STATES, RANK, STALE_MS, sessionLogOf, errorOf, asObject };++module.exports = { createRunTasks, presentResult, TERMINAL, STATE_LABEL, STATES, RANK, STALE_MS, sessionLogOf, errorOf, asObject, ATTEMPT_STATE, reasonOf, usageOf };@@@ -98,7 +98,7 @@ function createSessionLines({ store, config = {}, log = console }) {const retentionDays = Math.max(1, Number(config.sessions && config.sessions.retentionDays) || 90);/** Insert lines in order; an existing (task_id, seq) wins, so a replay never duplicates. → rows inserted. */-- async function append(taskId, lines) {++ async function append(taskId, lines, attemptN = null) {if (!taskId || !Array.isArray(lines) || !lines.length) return 0;const clean = [];for (const line of lines) {@@@ -114,12 +114,12 @@ function createSessionLines({ store, config = {}, log = console }) {const values = [];const params = [];chunk.forEach((line, j) => {-- const b = j * 5;-- values.push(`($${b + 1}, $${b + 2}, $${b + 3}::timestamptz, $${b + 4}, $${b + 5}::jsonb)`);-- params.push(taskId, line.seq, line.at, line.kind, JSON.stringify(line.body));++ const b = j * 6;++ values.push(`($${b + 1}, $${b + 2}, $${b + 3}::timestamptz, $${b + 4}, $${b + 5}::jsonb, $${b + 6})`);++ params.push(taskId, line.seq, line.at, line.kind, JSON.stringify(line.body), attemptN);});inserted += await db.exec(-- `INSERT INTO codes_session_lines (task_id, seq, at, kind, body) VALUES ${values.join(', ')}++ `INSERT INTO codes_session_lines (task_id, seq, at, kind, body, attempt_n) VALUES ${values.join(', ')}ON CONFLICT (task_id, seq) DO NOTHING`, params);}return inserted;@@@ -127,7 +127,7 @@ function createSessionLines({ store, config = {}, log = console }) {/** All rows of a task in stream order. */async function allRows(taskId) {-- return await db.many('SELECT seq, at, kind, body FROM codes_session_lines WHERE task_id = $1 ORDER BY seq ASC', [taskId]);++ return await db.many('SELECT seq, at, kind, body, attempt_n FROM codes_session_lines WHERE task_id = $1 ORDER BY seq ASC', [taskId]);}/**@@@ -135,7 +135,7 @@ function createSessionLines({ store, config = {}, log = console }) {* into the job's JSON lines; state and end frames become their own lines. The stream's seq is the base, so the* same event replayed after a reconnect maps to the same rows.*/-- function recorder(taskId) {++ function recorder(taskId, { attemptN = null } = {}) {let frameBuffer = '';let outputBuffer = '';let fallbackSeq = 0;@@@ -196,7 +196,7 @@ function createSessionLines({ store, config = {}, log = console }) {frameBuffer = `${frameBuffer}${String(text == null ? '' : text)}`.replace(/\r\n/g, '\n');const out = [];drainFrames(out);-- if (out.length) await append(taskId, out);++ if (out.length) await append(taskId, out, attemptN);},async flush() {const out = [];@@@ -208,7 +208,7 @@ function createSessionLines({ store, config = {}, log = console }) {out.push({ seq: (lastBase || 0) + PARTIAL_LINE_INDEX, at: store.iso(), kind: line.kind, body: line.body });outputBuffer = '';}-- if (out.length) await append(taskId, out);++ if (out.length) await append(taskId, out, attemptN);},ended: () => ended,};@@@ -219,7 +219,7 @@ function createSessionLines({ store, config = {}, log = console }) {* JSON, so a reconnect or a second result never adds a line twice; the new seqs sit after the streamed output and* before the terminal state/end line.*/-- async function backfill(row, result) {++ async function backfill(row, result, attemptN = null) {if (!row || !row.id) return 0;const rows = await allRows(row.id);const harness = rows.filter((r) => r && r.body && typeof r.body === 'object' && !Array.isArray(r.body)@@@ -242,7 +242,7 @@ function createSessionLines({ store, config = {}, log = console }) {const at = atIso(row.finished_at) || (row.updated_at ? iso(row.updated_at) : store.iso());return await append(row.id, missing.map((body, i) => ({seq: start + i, at, kind: typeof body.t === 'string' ? body.t : 'text', body,-- })));++ })), attemptN);}async function complete(taskId) {@@@ -252,8 +252,8 @@ function createSessionLines({ store, config = {}, log = console }) {);}-- async function backfillAndComplete(row, result) {-- const inserted = await backfill(row, result);++ async function backfillAndComplete(row, result, attemptN = null) {++ const inserted = await backfill(row, result, attemptN);await complete(row.id);return inserted;}@@@ -276,11 +276,11 @@ function createSessionLines({ store, config = {}, log = console }) {let source;if (rows.length) {source = 'session';-- lines = rows.filter((r) => Number(r.seq) > from).map((r) => ({ seq: Number(r.seq), at: r.at, kind: r.kind, body: r.body }));++ lines = rows.filter((r) => Number(r.seq) > from).map((r) => ({ seq: Number(r.seq), at: r.at, kind: r.kind, body: r.body, attempt_n: r.attempt_n == null ? null : Number(r.attempt_n) }));} else {source = 'result.log';const at = row.created_at ? iso(row.created_at) : store.iso();-- lines = (sessionLogOf(row) || []).map((body, i) => ({ seq: i + 1, at, kind: typeof body.t === 'string' ? body.t : 'text', body }))++ lines = (sessionLogOf(row) || []).map((body, i) => ({ seq: i + 1, at, kind: typeof body.t === 'string' ? body.t : 'text', body, attempt_n: 1 })).filter((line) => line.seq > from);}const page = lines.slice(0, max);@@@ -77,7 +77,7 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountSlet job = null;if (TERMINAL.has(state)) job = await run.read(jobId, tasks.who(row));-- return await store.tx(async () => {++ const applied = await store.tx(async () => {const inserted = await db.maybe(`INSERT INTO codes_events_inbox (event_id, event_type, task_id, outcome, processed_at)VALUES ($1, $2, $3, $4, $5::timestamptz) ON CONFLICT (event_id) DO NOTHING RETURNING event_id`,@@@ -90,6 +90,13 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountSawait db.exec('UPDATE codes_events_inbox SET outcome = $2 WHERE event_id = $1', [event.event_id, state]);return { duplicate: false, outcome: state, task_id: row.id };});++ // The attempt ended: the engine decides what happens next. It runs outside the transaction, because a++ // follow-up starts a Run job. A failure here is logged, not fatal: the delivery is applied, and the decision++ // is made again by the fallback (the read of Run after 10 quiet minutes).++ if (!applied.duplicate) {++ try { await tasks.afterEnd(applied.task_id); } catch (err) { log.warn(`[Codes] task ${applied.task_id} decision failed: ${err.message}`); }++ }++ return applied;}const router = express.Router();@@@ -5,7 +5,8 @@* GET /run the form, the person's projects and their recent tasks (a sign-in prompt for a guest)* POST /run start one: a run_tasks row and an OpenVibe Run job for the person (server/domain/run-client.js)* POST /run/project create a project for the runs when the person has none they may run jobs in-- * GET /run/tasks/:id one task: live while it runs, then the agents' summary, the changed files and the patch++ * GET /run/tasks/:id one task: its attempts (the timeline), live while it runs, then the agents' summary,++ * the changed files and the patch* GET /run/tasks/:id/events its live lines, proxied from Run's job stream (server-sent events)* POST /run/tasks/:id/cancel stop it*@@@ -26,6 +27,9 @@ const DELEGABLE = new Set(['owner', 'admin']);const MACHINES_URL = 'https://openvibe.bot/machines';const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;++/** The task's outcome in the words a person reads. */++const OUTCOME_TEXT = { done: 'Done', unfinished: "Didn't finish", stopped: 'Stopped by you' };++/** A repository URL as a heading: the host small, then owner/name, which may break after a slash but never mid-word. */function repoName(url) {const [host, ...path] = String(url).replace(/^https:\/\//, '').split('/');@@@ -182,6 +186,24 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat</section>`;}++ /** The task's timeline: every attempt and why it happened. The data-* hooks are for the design of stage C4. */++ async function timelineHtml(row) {++ const attempts = await taskStore.attempts.list(row.id);++ if (!attempts.length) return '';++ const outcome = row.outcome++ ? html`<p class="run-outcome" data-outcome="${row.outcome}"><b>${OUTCOME_TEXT[row.outcome] || row.outcome}</b>${row.outcome_note ? html` ${row.outcome_note}` : ''}</p>`++ : '';++ return html`<section class="run-timeline" data-timeline data-attempts="${attempts.length}">++<h2>The attempts</h2>++<ol class="run-attempts">${attempts.map((a) => html`<li class="run-attempt" data-attempt="${a.n}" data-attempt-id="${a.id}" data-state="${a.state}" data-reason="${a.reason || ''}" data-kind="${a.kind || 'first'}" data-agent="${a.agent || ''}">++<p class="run-attempt-head"><b>Attempt ${a.n}</b>${a.agent ? html` · ${a.agent}` : ''} <span class="run-pill run-pill-${a.state}">${a.state}</span></p>++<p class="run-attempt-why" data-why>${a.why || ''}</p>++${a.reason ? html`<p class="run-attempt-end" data-end-why>${taskStore.attempts.wordsOf(a.reason)}</p>` : ''}++</li>`)}</ol>++${outcome}++</section>`;++ }++r.get('/run/tasks/:id', async (req, res) => {if (!signedIn(req)) return res.redirect(302, `/auth/login?next=${encodeURIComponent(req.originalUrl)}`);let row = await own(req, req.params.id);@@@ -197,6 +219,7 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat<h1 class="run-repo">${repoName(row.repo)}${row.ref ? html` <small>${row.ref}</small>` : ''}</h1><p><span class="run-pill run-pill-${row.state}" data-state>${STATE_LABEL[row.state] || row.state}</span>${row.agent ? html` <span class="run-pill">${row.agent}</span>` : ''}</p><pre tabindex="0" class="run-prompt">${row.prompt}</pre>++${await timelineHtml(row)}${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events" data-title="${title}" aria-live="polite"><h2>Live</h2><ol class="run-log" data-log></ol></section><form method="post" action="/run/tasks/${row.id}/cancel"><button class="run-btn" type="submit">Cancel</button></form>` : html`${session}${resultHtml(row, { session: !!(lines && lines.length) })}`}`;return page(req, res, { title: 'Task', body, sharedScripts: ['agent-session.js'] });@@@ -242,7 +265,9 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/eventsif (!resp.ok || !resp.body) return res.status(resp.status === 404 ? 404 : 503).end();res.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-store', 'X-Accel-Buffering': 'no', Connection: 'keep-alive' });// The bytes go to the browser unchanged, and the same frames become codes_session_lines as they stream.-- const recorder = sessionStore.recorder(row.id);++ const attempts = await taskStore.attempts.list(row.id);++ const attemptN = attempts.length ? Number(attempts[attempts.length - 1].n) : null;++ const recorder = sessionStore.recorder(row.id, { attemptN });const decoder = new TextDecoder();try {for await (const chunk of resp.body) {@@@ -286,13 +311,8 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/eventsif (!sameOrigin(req)) return res.status(403).type('text/plain').send('Cross-site form posts are refused.\n');const row = await own(req, req.params.id);if (!row) return res.status(404).type('text/plain').send('No such task of yours.\n');-- if (row.job_id && !TERMINAL.has(row.state)) {-- try { await run.cancel(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }-- try {-- const fresh = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [row.id]);-- if (fresh) await taskStore.refresh(fresh, { force: true });-- } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }-- }++ // The running attempt stops: the outcome is "stopped" and nothing carries it on.++ try { await taskStore.stopTask(row); } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }return res.redirect(303, `/run/tasks/${row.id}`);});nnew file mode 100644@@@ -0,0 +1,151 @@++'use strict';++++/**++ * A task's event stream (GET /api/v1/tasks/:id/events): the session's lines in the order they happened, the attempt++ * each line belongs to, and the attempts themselves (started, ended with why). The stream is resumable: a client++ * that reconnects sends the last line's seq as Last-Event-ID and gets only what came after it. A heartbeat every 15++ * seconds keeps the connection alive through a proxy that would otherwise drop it.++ *++ * createTaskEvents({ store, config, log, attempts })++ * stream(req, res, row) the stream of one task++ *++ * The lines come from codes_session_lines (the session the machine streamed, back-filled from its final result) and++ * the attempts from codes_attempts. Nothing here writes anything: the stream only reads.++ */++++const { TERMINAL } = require('../domain/run-tasks');++++const POLL_MS = 1000;++const PAGE = 500;++const ENDED = ['done', 'failed', 'stopped', 'budget-stopped', 'interrupted'];++++function createTaskEvents({ store, config = {}, log = console, attempts = null }) {++ const db = store.db;++ const heartbeatMs = Math.max(1000, Number(config.sse && config.sse.heartbeatMs) || 15_000);++ const attemptStore = attempts || require('../domain/attempts').createAttempts({ store, config, log });++++ /** One SSE frame. */++ const frame = (event, data, id = null) => (id == null ? '' : `id: ${id}\n`) + `event: ${event}\ndata: ${JSON.stringify(data)}\n\n`;++++ const started = (a) => ({++ type: 'attempt_started', attempt_n: Number(a.n), id: a.id, agent: a.agent || null, model: a.model || null,++ why: a.why || null, state: 'queued', at: a.created_at ? new Date(a.created_at).toISOString() : null,++ });++ const ended = (a) => ({++ type: 'attempt_ended', attempt_n: Number(a.n), id: a.id, agent: a.agent || null, model: a.model || null,++ why: a.why || null, state: a.state, reason: a.reason || null,++ end_why: a.reason ? attemptStore.wordsOf(a.reason) : null,++ cost_usd: a.cost_usd == null ? null : Number(a.cost_usd),++ tokens: a.tokens == null ? null : Number(a.tokens),++ steps: a.steps == null ? null : Number(a.steps),++ at: a.finished_at ? new Date(a.finished_at).toISOString() : null,++ });++++ /** The stream of one task: its lines, its attempts, a heartbeat, and an end when the task has ended. */++ async function stream(req, res, row) {++ res.writeHead(200, {++ 'Content-Type': 'text/event-stream',++ 'Cache-Control': 'no-store',++ 'X-Accel-Buffering': 'no',++ Connection: 'keep-alive',++ });++ if (res.flushHeaders) res.flushHeaders();++++ const asked = String(req.get('last-event-id') || (req.query && req.query.last_event_id) || '0').trim();++ let lastSeq = /^\d+$/.test(asked) ? Number(asked) : 0;++ let closed = false;++ let busy = false;++ let terminalAt = 0;++++ const write = (text) => {++ if (closed || !text) return;++ res.write(text);++ };++++ // The attempts a client has not seen: from the attempt its last line belongs to.++ const anchor = lastSeq++ ? await db.maybe('SELECT attempt_n FROM codes_session_lines WHERE task_id = $1 AND seq = $2', [row.id, lastSeq])++ : null;++ const fromN = anchor && anchor.attempt_n ? Number(anchor.attempt_n) : 1;++ const known = await attemptStore.list(row.id);++ const sent = new Map(known.map((a) => [a.id, a.state]));++ let out = '';++ for (const a of known) {++ if (Number(a.n) < fromN) continue;++ out += frame('attempt', started(a));++ if (ENDED.includes(a.state)) out += frame('attempt', ended(a));++ }++ write(out);++++ const sendLines = async () => {++ const lines = await db.many(++ 'SELECT seq, at, kind, body, attempt_n FROM codes_session_lines WHERE task_id = $1 AND seq > $2 ORDER BY seq ASC LIMIT $3',++ [row.id, lastSeq, PAGE],++ );++ for (const line of lines) {++ write(frame('line', {++ seq: Number(line.seq),++ at: line.at ? new Date(line.at).toISOString() : null,++ kind: line.kind,++ attempt_n: line.attempt_n == null ? null : Number(line.attempt_n),++ body: line.body,++ }, Number(line.seq)));++ lastSeq = Math.max(lastSeq, Number(line.seq));++ }++ return lines.length;++ };++ await sendLines();++++ const tick = async () => {++ if (closed || busy) return;++ busy = true;++ try {++ const sentLines = await sendLines();++ const fresh = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [row.id]);++ const rows = await attemptStore.list(row.id);++ let updates = '';++ for (const a of rows) {++ const was = sent.get(a.id);++ if (was === a.state) continue;++ sent.set(a.id, a.state);++ if (was === undefined) updates += frame('attempt', started(a));++ if (ENDED.includes(a.state)) updates += frame('attempt', ended(a));++ }++ write(updates);++ if (fresh && TERMINAL.has(fresh.state) && !rows.some((a) => ['queued', 'running'].includes(a.state))) {++ if (!terminalAt) terminalAt = Date.now();++ // The task has ended and its outcome is known: the stream is done. A task whose decision is still++ // being made (no outcome yet) keeps the stream a little longer, so a follow-up is not missed.++ if (fresh.outcome || Date.now() - terminalAt > 10_000) {++ closed = true;++ clearInterval(poll);++ clearInterval(beat);++ res.end();++ }++ } else {++ terminalAt = 0;++ }++ } catch (e) {++ log.warn(`[Codes] task ${row.id} events: ${e.message}`);++ } finally {++ busy = false;++ }++ };++++ const poll = setInterval(tick, POLL_MS);++ const beat = setInterval(() => write(': heartbeat\n\n'), heartbeatMs);++ const stop = () => {++ if (closed) return;++ closed = true;++ clearInterval(poll);++ clearInterval(beat);++ };++ req.on('close', stop);++ res.on('close', stop);++ return undefined;++ }++++ return { stream, heartbeatMs };++}++++module.exports = { createTaskEvents };@@@ -7,19 +7,29 @@* machine, in the project of their latest /run task.** POST /tasks codes.task.start codes.task-start-request@1 → codes.task-start-result@1 (201, or 200 again)-- * GET /tasks/:id codes.task.read → codes.task@1++ * GET /tasks/:id codes.task.read → codes.task@2 (the task as codes.task@1, with its attempts and++ * its outcome added to it)* POST /tasks/:id/cancel codes.task.cancel → codes.task@1++ * GET /tasks/:id/events codes.task.read → the task's event stream (server-sent events)++ * POST /tasks/:id/continue codes.task.start → the next attempt now (a switch when an agent is named)++ * POST /tasks/:id/stop codes.task.cancel → the running attempt stops; nothing carries it on** A Network service token for audience openvibe.codes (CODES_AUDIENCE), from a svc: principal holding the route's-- * capability. A service sees only the tasks it started, for that person; anything else is 404. The person's task-- * page (the codes.task@1 url) shows every task they own, whoever started it. Calls count against the person's own-- * per-actor budget (actor-limits.js), not the service's.++ * capability. A service sees only the tasks it started, for that person; anything else is 404. The owner may read++ * and act on their own tasks without a token. The person's task page (the codes.task@1 url) shows every task they++ * own, whoever started it. Calls count against the person's own per-actor budget (actor-limits.js), not the service's.++ *++ * Stage C2: a task's timeline is its attempts. GET /tasks/:id answers codes.task@2: the task itself with the same++ * fields it always had, and, added to them, its attempts and its outcome in plain words ("done", "unfinished" with++ * what is left, or "stopped"). GET /tasks/:id/events replays the session's lines in order — resumable with++ * Last-Event-ID, a heartbeat every 15 s — with the attempt each line belongs to.*/const express = require('express');const { serviceAuth, capabilities, http, validate } = require('openvibe-contracts');const { asyncRouter } = require('./router');const { parseArgs } = require('../../harness/job');const { createRunTasks, TERMINAL } = require('../domain/run-tasks');++const { createTaskEvents } = require('./task-events');const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;const TASK = /^tsk_[0-9A-HJKMNP-TV-Z]{26}$/;@@@ -28,6 +38,7 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, taskconst r = asyncRouter();const db = store.db;const taskStore = tasks || createRunTasks({ store, run, config, log });++ const taskEvents = createTaskEvents({ store, config, log, attempts: taskStore.attempts });const problem = (req, res, status, code, detail) => http.sendProblem(res, status, code, { detail, ctx: req.ov });const json = express.json({ limit: '32kb' });@@@ -53,10 +64,32 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task/** The task this service started for this person, or null. */const mine = async (req) => {if (!TASK.test(String(req.params.id))) return null;++ if (req.principal.owner) return db.maybe('SELECT * FROM run_tasks WHERE id = $1 AND subject = $2', [req.params.id, req.principal.subject]);return db.maybe('SELECT * FROM run_tasks WHERE id = $1 AND subject = $2 AND via = $3', [req.params.id, req.principal.subject, req.principal.via]);};const byKey = (p, key) => db.maybe('SELECT * FROM run_tasks WHERE subject = $1 AND via = $2 AND idempotency_key = $3', [p.subject, p.via, key]);++ /** The owner (signed in), or a service token with the capability acting for the person in X-OV-Subject. */++ const signedIn = (req) => req.viewer && req.viewer.kind === 'user' && req.viewer.subject;++ const guardAny = (cap) => async (req, res, next) => {++ if (String(req.headers.authorization || '').startsWith('Bearer ')) return guard(cap)(req, res, next);++ if (signedIn(req)) {++ req.principal = { sub: `user:${req.viewer.subject}`, subject: req.viewer.subject, via: null, owner: true };++ return next();++ }++ return problem(req, res, 401, 'token.required', `${cap} needs a Network service token for ${config.serviceAudience}, or a signed-in owner`);++ };++++ /** The task, its attempts and its outcome in plain words. The task's own fields stay at the top level. */++ const timeline = async (row) => {++ const attempts = await taskStore.attempts.list(row.id);++ return {++ ...taskStore.present(row),++ attempts: attempts.map((a) => taskStore.attempts.present(a)),++ outcome: { state: row.outcome || null, note: row.outcome_note || null },++ };++ };++r.post('/tasks', guard('codes.task.start'), actorLimits.budget('codes.task.start'), json, async (req, res) => {const b = req.body;// Actor back-reference: the contract does not carry these yet, so they are stripped before validation and stored@@@ -102,10 +135,41 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, taskreturn res.status(201).json({ task: taskStore.present(row), created: true });});-- r.get('/tasks/:id', guard('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {++ r.get('/tasks/:id', guardAny('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {++ const row = await mine(req);++ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');++ return res.json(await timeline(await taskStore.refresh(row)));++ });++++ /** The task's event stream: the session's lines in order, resumable with Last-Event-ID, plus the attempts. */++ r.get('/tasks/:id/events', guardAny('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {++ const row = await mine(req);++ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');++ return await taskEvents.stream(req, res, row);++ });++++ /** The next attempt now: the owner carries the task on, or hands it to another agent. */++ r.post('/tasks/:id/continue', guardAny('codes.task.start'), actorLimits.budget('codes.task.start'), json, async (req, res) => {++ const row = await mine(req);++ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');++ const b = req.body || {};++ const note = typeof b.note === 'string' ? b.note.slice(0, 4000) : null;++ const agent = typeof b.agent === 'string' && b.agent ? b.agent : null;++ if (agent && !harnesses.list().some((a) => a.id === agent)) return problem(req, res, 422, 'codes.agent_unknown', 'agent is one of GET /api/v1/harnesses');++ const out = await taskStore.continueTask(row, { note, agent });++ if (out.error) return problem(req, res, 409, 'codes.task_conflict', out.error);++ log.info(`[Codes] task ${row.id} carried on (${agent ? `with ${agent}` : 'by the owner'})`);++ return res.json(await timeline(out.row));++ });++++ /** The running attempt stops: the outcome is "stopped" and nothing carries it on. */++ r.post('/tasks/:id/stop', guardAny('codes.task.cancel'), actorLimits.budget('codes.task.start'), async (req, res) => {const row = await mine(req);if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');-- return res.json(taskStore.present(await taskStore.refresh(row)));++ const out = await taskStore.stopTask(row);++ if (out.error) return problem(req, res, 409, 'codes.task_conflict', out.error);++ log.info(`[Codes] task ${row.id} stopped by ${req.principal.via || req.principal.subject}`);++ return res.json(await timeline(out.row));});r.post('/tasks/:id/cancel', guard('codes.task.cancel'), actorLimits.budget('codes.task.start'), async (req, res) => {nnew file mode 100644@@@ -0,0 +1,288 @@++'use strict';++++/**++ * Stage C2: a task's attempts. Every Run job is one attempt; when an attempt ends the engine decides what happens++ * next — done, a continue in the same conversation, a switch to another agent, a wrap-up after a limit, or an++ * unfinished task — and one ending starts at most one follow-up (the portal's 10-30 bug, where two racing ticks++ * carried one job on twice).++ *++ * The migration (an existing task becomes attempt 1), every engine decision, the fence under two concurrent endings,++ * the stop fence, and the auth of every new route.++ */++const assert = require('assert');++const { boot, check, done } = require('./helpers/boot');++++const TASK = /^tsk_[0-9A-HJKMNP-TV-Z]{26}$/;++++(async () => {++ const calls = [];++ const jobs = new Map();++ let counter = 0;++ const run = {++ enabled: true,++ async submit(o) {++ counter++;++ const id = `job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J${String(counter).padStart(2, '0')}`;++ jobs.set(id, { id, state: 'queued', args: o.args });++ calls.push(['submit', o]);++ return { id, state: 'queued' };++ },++ async read(jobId) { return jobs.get(jobId) || { id: jobId, state: 'queued' }; },++ async cancel(jobId, who) { calls.push(['cancel', jobId, who]); const j = jobs.get(jobId); if (j) j.state = 'cancelled'; return j; },++ async streamUrl() { throw new Error('not used'); },++ };++ const t = await boot({ run, env: { CODES_TASK_STEPS: '2' } });++ const db = t.ctx.store.db;++ const ada = t.network.addUser('ada');++ const bo = t.network.addUser('bo');++ const CAPS = ['codes.task.start', 'codes.task.read', 'codes.task.cancel'];++ const actor = t.network.mintService('svc:actor', 'openvibe.codes', CAPS);++ const project = 'prj_01J8Z3Q4R5S6T7V8W9X0Y1Z2P1';++++ const call = async (method, path, { token = actor, subject = ada.subject, body, headers = {} } = {}) => {++ const h = { Accept: 'application/json', ...headers, ...(token ? { Authorization: `Bearer ${token}` } : {}), ...(subject ? { 'X-OV-Subject': subject } : {}) };++ if (body !== undefined) h['Content-Type'] = 'application/json';++ const res = await fetch(`${t.base}/api/v1${path}`, { method, headers: h, body: body === undefined ? undefined : JSON.stringify(body) });++ return { status: res.status, json: await res.json().catch(() => null) };++ };++++ /** A task for Ada, started through the API (so it has its first attempt). */++ const startTask = async (prompt = 'Fix the parser and add a test for it.') => {++ const r = await call('POST', '/tasks', { body: { repo: 'https://github.com/octocat/Hello-World', prompt, idempotency_key: `act_${Math.random().toString(36).slice(2, 10)}` } });++ assert.strictEqual(r.status, 201, JSON.stringify(r.json));++ return r.json.task.id;++ };++++ /** The attempt ends: the state change and then the engine's decision (what the events consumer does). */++ const endAttempt = async (taskId, { state = 'failed', result = null, error = null } = {}) => {++ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [taskId]);++ const job = { id: row.job_id, state, result, error: error ? { code: 'run.job.worker_failed', detail: error } : null };++ const updated = await t.ctx.tasks.applyEvent(row, { state, job, error });++ await t.ctx.tasks.afterEnd(updated.id);++ return await rowOf(updated.id);++ };++++ const attemptsOf = async (taskId) => await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);++ const rowOf = async (taskId) => await db.one('SELECT * FROM run_tasks WHERE id = $1', [taskId]);++++ try {++ await check('the migration: every task that exists becomes attempt 1 of itself', async () => {++ // A task from before the attempts (written straight into the table, as an old row would be).++ const old = 'tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T9';++ await db.exec(`INSERT INTO run_tasks (id, subject, project_id, job_id, repo, prompt, agent, state, created_at, updated_at)++ VALUES ($1, $2, $3, 'job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J99', 'https://github.com/a/b', 'by hand', 'claude-code', 'succeeded', 1, 2)`, [old, ada.subject, project]);++ const read = await t.ctx.attempts.list(old);++ assert.strictEqual(read.length, 1, 'one attempt');++ assert.strictEqual(Number(read[0].n), 1);++ assert.strictEqual(read[0].state, 'done', 'the task succeeded, so its first attempt did');++ assert.ok(read[0].why, 'the attempt says why it happened');++ const shown = await t.get(`/api/v1/tasks/${old}`, { as: ada });++ assert.strictEqual(shown.status, 200, shown.text.slice(0, 200));++ assert.strictEqual(shown.json().attempts.length, 1);++ assert.strictEqual(shown.json().attempts[0].n, 1);++ });++++ await check('a task starts as one attempt, and the timeline says why', async () => {++ const id = await startTask();++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 1);++ assert.strictEqual(Number(attempts[0].n), 1);++ assert.strictEqual(attempts[0].state, 'queued');++ assert.strictEqual(attempts[0].kind, 'first');++ assert.ok(attempts[0].job_id, 'the attempt carries the Run job');++ const read = await call('GET', `/tasks/${id}`);++ assert.strictEqual(read.json.id, id, 'the task at the top level, as codes.task@1 had it');++ assert.strictEqual(read.json.state, 'queued');++ assert.ok(read.json.url, 'the task carries its own fields');++ assert.strictEqual(read.json.attempts.length, 1);++ assert.strictEqual(read.json.outcome.state, null, 'a running task has no outcome yet');++ });++++ await check('a finished attempt: the task is done, and nothing follows it', async () => {++ const id = await startTask();++ const updated = await endAttempt(id, { state: 'succeeded', result: { ok: true, summary: 'all done', files: [], log: [{ t: 'done', ok: true, text: 'all done' }] } });++ assert.strictEqual(updated.outcome, 'done');++ assert.strictEqual((await attemptsOf(id)).length, 1, 'no follow-up');++ const read = await call('GET', `/tasks/${id}`);++ assert.strictEqual(read.json.outcome.state, 'done');++ assert.strictEqual(read.json.attempts[0].state, 'done');++ });++++ await check('a crash: the same conversation continues, with the session to resume', async () => {++ const id = await startTask();++ const result = {++ ok: false, summary: 'I will continue with the rest of it', files: [],++ log: [{ t: 'agent', agent: 'claude-code', session: 'sess-42' }, { t: 'done', ok: false, text: 'I will continue with the rest of it' }],++ };++ await endAttempt(id, { state: 'failed', result, error: 'the machine went away' });++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 2, 'the task was carried on');++ assert.strictEqual(attempts[1].kind, 'continue');++ assert.strictEqual(attempts[1].session, 'sess-42', 'the session to resume');++ assert.ok(/^carried on in the same conversation: /.test(attempts[1].why), attempts[1].why);++ // The new attempt is a Run job of its own, with RESUME set to the session.++ const submits = calls.filter((c) => c[0] === 'submit');++ const last = submits[submits.length - 1][1];++ assert.strictEqual(last.args.resume, 'sess-42');++ assert.ok(last.args.prompt.includes('Continuation'), 'the prompt tells the agent to carry on');++ assert.notStrictEqual(last.key, id, 'the follow-up has its own idempotency key');++ });++++ await check('a limit: another agent takes over, with a hand-over note of what the last one did', async () => {++ const id = await startTask();++ const result = { ok: false, summary: 'I fixed the parser', files: [{ path: 'src/parse.js', added: 3, removed: 1 }], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] };++ await endAttempt(id, { state: 'failed', result, error: 'usage limit reached' });++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 2);++ assert.strictEqual(attempts[1].kind, 'switch');++ assert.ok(attempts[1].agent && attempts[1].agent !== attempts[0].agent, `${attempts[1].agent} took over from ${attempts[0].agent}`);++ assert.ok(/took over: /.test(attempts[1].why), attempts[1].why);++ const submits = calls.filter((c) => c[0] === 'submit');++ const last = submits[submits.length - 1][1];++ assert.ok(last.args.prompt.includes('Hand-over'), 'the new agent gets a hand-over note');++ assert.ok(last.args.prompt.includes('src/parse.js'), 'the note says what changed');++ assert.ok(last.args.prompt.includes('I fixed the parser'), 'the note says what it said');++ });++++ await check('a limit of the owner: one last turn to land the work', async () => {++ const id = await startTask();++ const result = {++ ok: false, summary: 'half done', files: [],++ log: [{ t: 'tool', id: 't1', name: 'Edit' }, { t: 'tool', id: 't2', name: 'Edit' }, { t: 'done', ok: false, text: 'maximum turns' }],++ };++ await endAttempt(id, { state: 'failed', result, error: 'the agent stopped' });++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 2);++ assert.strictEqual(attempts[0].state, 'budget-stopped', 'a limit is not a failure');++ assert.strictEqual(attempts[0].reason, 'budget');++ assert.strictEqual(attempts[1].kind, 'wrap-up');++ assert.ok(/one last turn to land the work/.test(attempts[1].why), attempts[1].why);++ const submits = calls.filter((c) => c[0] === 'submit');++ assert.ok(submits[submits.length - 1][1].args.prompt.includes('one short last turn'), 'the wrap-up prompt');++ });++++ await check('the same failure twice: the task is unfinished, with what is left', async () => {++ const id = await startTask();++ const fail = { ok: false, summary: 'it broke', files: [], log: [{ t: 'done', ok: false, text: 'the tool call failed' }] };++ await endAttempt(id, { state: 'failed', result: fail, error: 'the tool call failed' });++ await endAttempt(id, { state: 'failed', result: fail, error: 'the tool call failed' });++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 2, 'a second opinion, and no more');++ const row = await rowOf(id);++ assert.strictEqual(row.outcome, 'unfinished');++ assert.ok(row.outcome_note && row.outcome_note.length > 10, 'the outcome says what is left');++ });++++ await check('one ending starts at most one follow-up, even when two endings race', async () => {++ const id = await startTask();++ const before = calls.filter((c) => c[0] === 'submit').length;++ const row = await rowOf(id);++ const job = { id: row.job_id, state: 'failed', result: { ok: false, summary: 'x', files: [], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] }, error: { code: 'run.job.worker_failed', detail: 'usage limit reached' } };++ // Two ticks see the same ending at the same time (the portal's 10-30 bug).++ await Promise.all([++ t.ctx.tasks.applyEvent(row, { state: 'failed', job, error: 'usage limit reached' }),++ t.ctx.tasks.applyEvent(row, { state: 'failed', job, error: 'usage limit reached' }),++ ]);++ await Promise.all([t.ctx.tasks.afterEnd(id), t.ctx.tasks.afterEnd(id)]);++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.length, 2, 'one follow-up, not two');++ const submits = calls.filter((c) => c[0] === 'submit').length - before;++ assert.strictEqual(submits, 1, 'one job started, not two');++ });++++ await check('a stop fences the task: the outcome is stopped and nothing carries it on', async () => {++ const id = await startTask();++ const before = calls.filter((c) => c[0] === 'submit').length;++ const stopped = await call('POST', `/tasks/${id}/stop`);++ assert.strictEqual(stopped.status, 200, JSON.stringify(stopped.json));++ assert.strictEqual(stopped.json.outcome.state, 'stopped');++ assert.ok(calls.some((c) => c[0] === 'cancel'), 'the job was cancelled');++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts[0].state, 'stopped');++ assert.strictEqual(attempts[0].reason, 'stopped');++ // The machine's late report changes nothing.++ await endAttempt(id, { state: 'failed', error: 'the machine went away' });++ assert.strictEqual((await attemptsOf(id)).length, 1, 'nothing carried it on');++ assert.strictEqual(calls.filter((c) => c[0] === 'submit').length - before, 0, 'no job was started');++ assert.strictEqual((await rowOf(id)).outcome, 'stopped');++ });++++ await check('the owner carries the task on, or hands it to another agent', async () => {++ const id = await startTask();++ await endAttempt(id, { state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] } });++ const continued = await call('POST', `/tasks/${id}/continue`, { body: { note: 'Only the tests are left.' } });++ assert.strictEqual(continued.status, 200, JSON.stringify(continued.json));++ assert.strictEqual(continued.json.attempts.length, 2);++ assert.strictEqual(continued.json.attempts[1].kind, 'manual');++ assert.ok(/carried on by the owner/.test(continued.json.attempts[1].why), continued.json.attempts[1].why);++ // While the task runs again, nobody carries it on a second time.++ const again = await call('POST', `/tasks/${id}/continue`, { body: { agent: 'codex' } });++ assert.strictEqual(again.status, 409, 'the task is running again');++ });++++ await check('the auth of every new route: no token, no capability, another person', async () => {++ const id = await startTask();++ const noCap = t.network.mintService('svc:nocap', 'openvibe.codes', ['codes.other']);++ const other = t.network.mintService('svc:other', 'openvibe.codes', CAPS);++ // No token at all.++ for (const [method, path] of [['GET', `/tasks/${id}`], ['GET', `/tasks/${id}/events`], ['POST', `/tasks/${id}/continue`], ['POST', `/tasks/${id}/stop`]]) {++ const res = await call(method, path, { token: null });++ assert.strictEqual(res.status, 401, `${method} ${path} without a token`);++ }++ // A service without the capability.++ assert.strictEqual((await call('GET', `/tasks/${id}`, { token: noCap })).status, 403);++ assert.strictEqual((await call('POST', `/tasks/${id}/continue`, { token: noCap, body: {} })).status, 403);++ // Another person, and another service (which sees only the tasks it started).++ assert.strictEqual((await call('GET', `/tasks/${id}`, { subject: bo.subject })).status, 404);++ assert.strictEqual((await call('GET', `/tasks/${id}`, { token: other })).status, 404);++ assert.strictEqual((await call('POST', `/tasks/${id}/stop`, { subject: bo.subject })).status, 404);++ // A task that does not exist.++ assert.strictEqual((await call('GET', '/tasks/tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T1')).status, 404);++ assert.strictEqual((await call('GET', '/tasks/nope')).status, 404);++ });++++ await check('the task page shows the timeline of attempts, each with its why', async () => {++ const id = await startTask();++ const result = { ok: false, summary: 'x', files: [], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] };++ await endAttempt(id, { state: 'failed', result, error: 'usage limit reached' });++ const page = await t.get(`/run/tasks/${id}`, { as: ada });++ assert.strictEqual(page.status, 200, page.text.slice(0, 300));++ assert.ok(page.text.includes('data-timeline'), 'the timeline is there');++ assert.ok(page.text.includes('data-attempt="1"') && page.text.includes('data-attempt="2"'), 'every attempt is a row');++ assert.ok(page.text.includes('data-state="failed"'), 'the attempt says how it ended');++ assert.ok(page.text.includes('data-why'), 'the attempt says why it happened');++ assert.ok(/took over: /.test(page.text), 'the second attempt says who took over');++ assert.ok(page.text.indexOf('data-timeline') < page.text.indexOf('data-session') || !page.text.includes('data-session'),++ 'the timeline is above the session');++ });++++ await check('a task is carried on twice, and then it needs a person', async () => {++ const id = await startTask();++ const crash = {++ ok: false, summary: 'I will continue with the rest of it', files: [],++ log: [{ t: 'agent', agent: 'claude-code', session: 'sess-1' }, { t: 'done', ok: false, text: 'I will continue with the rest of it' }],++ };++ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });++ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });++ const attempts = await attemptsOf(id);++ assert.strictEqual(attempts.filter((a) => a.kind === 'continue').length, 2, 'two automatic continues');++ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });++ const row = await rowOf(id);++ assert.strictEqual(row.outcome, 'unfinished', 'the third one needs a person');++ assert.ok(/carried on 2 times/.test(row.outcome_note), row.outcome_note);++ assert.strictEqual((await attemptsOf(id)).length, 3, 'no third continue');++ });++++ await check('the owner needs no token: their own task reads, continues and stops', async () => {++ const id = await startTask();++ const mine = await t.get(`/api/v1/tasks/${id}`, { as: ada });++ assert.strictEqual(mine.status, 200, mine.text.slice(0, 200));++ assert.strictEqual(mine.json().attempts.length, 1);++ const theirs = await t.get(`/api/v1/tasks/${id}`, { as: bo });++ assert.strictEqual(theirs.status, 404);++ });++ } finally {++ await t.close();++ }++ done();++})().catch((e) => { console.error(e); process.exit(1); });nnew file mode 100644@@@ -0,0 +1,138 @@++'use strict';++++/**++ * A task's event stream (GET /api/v1/tasks/:id/events): the session's lines in the order they happened, the attempt++ * each line belongs to, and the attempts themselves (started, ended with why). A client that reconnects sends the last++ * line's seq as Last-Event-ID and gets only what came after it; a heartbeat keeps the connection alive.++ */++const assert = require('assert');++const http = require('http');++const { boot, check, done } = require('./helpers/boot');++++(async () => {++ const calls = [];++ let counter = 0;++ const run = {++ enabled: true,++ async submit(o) {++ counter++;++ const id = `job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J${String(counter).padStart(2, '0')}`;++ calls.push(['submit', o]);++ return { id, state: 'queued' };++ },++ async read(jobId) { return { id: jobId, state: 'queued' }; },++ async cancel(jobId) { return { id: jobId, state: 'cancelled' }; },++ async streamUrl() { throw new Error('not used'); },++ };++ const t = await boot({ run, env: { CODES_SSE_HEARTBEAT_MS: '300' } });++ const db = t.ctx.store.db;++ const ada = t.network.addUser('ada');++ const bo = t.network.addUser('bo');++ const CAPS = ['codes.task.start', 'codes.task.read', 'codes.task.cancel'];++ const actor = t.network.mintService('svc:actor', 'openvibe.codes', CAPS);++++ const call = async (method, path, { token = actor, subject = ada.subject, body, headers = {} } = {}) => {++ const h = { Accept: 'text/event-stream', ...headers, ...(token ? { Authorization: `Bearer ${token}` } : {}), ...(subject ? { 'X-OV-Subject': subject } : {}) };++ if (body !== undefined) h['Content-Type'] = 'application/json';++ const res = await fetch(`${t.base}/api/v1${path}`, { method, headers: h, body: body === undefined ? undefined : JSON.stringify(body) });++ return { status: res.status, text: await res.text() };++ };++++ /** The stream, read until it ends (a finished task) or until the marker arrives. */++ const stream = (path, { headers = {}, until = null, ms = 8000 } = {}) => new Promise((resolve, reject) => {++ const req = http.get(`${t.base}${path}`, { headers }, (res) => {++ let text = '';++ res.setEncoding('utf8');++ const timer = setTimeout(() => { req.destroy(); resolve({ status: res.statusCode, text }); }, ms);++ res.on('data', (chunk) => {++ text += chunk;++ if (until && text.includes(until)) { clearTimeout(timer); req.destroy(); resolve({ status: res.statusCode, text }); }++ });++ res.on('end', () => { clearTimeout(timer); resolve({ status: res.statusCode, text }); });++ });++ req.on('error', (e) => (e.code === 'ECONNRESET' ? resolve({ status: 0, text: '' }) : reject(e)));++ });++++ /** A task of Ada's with two attempts and three lines of session. */++ const makeTask = async () => {++ const r = await fetch(`${t.base}/api/v1/tasks`, {++ method: 'POST',++ headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${actor}`, 'X-OV-Subject': ada.subject },++ body: JSON.stringify({ repo: 'https://github.com/octocat/Hello-World', prompt: 'Add a test for the date parser.', idempotency_key: `act_${Math.random().toString(36).slice(2, 10)}` }),++ });++ const json = await r.json();++ const id = json.task.id;++ await db.exec(`INSERT INTO codes_session_lines (task_id, seq, at, kind, body, attempt_n) VALUES++ ($1, 1, now(), 'text', '{"t":"text","text":"first"}'::jsonb, 1),++ ($1, 2, now(), 'text', '{"t":"text","text":"second"}'::jsonb, 1),++ ($1, 3, now(), 'text', '{"t":"text","text":"third"}'::jsonb, 2)`, [id]);++ return id;++ };++++ try {++ // Ada ran a task on /run before (so Codes knows her project).++ await db.exec(`INSERT INTO run_tasks (id, subject, project_id, job_id, repo, prompt, state, created_at, updated_at)++ VALUES ('tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T1', $1, 'prj_01J8Z3Q4R5S6T7V8W9X0Y1Z2P1', 'job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J00', 'https://github.com/a/b', 'by hand', 'succeeded', 1, 2)`, [ada.subject]);++++ await check('the stream: the lines in order, the attempt each belongs to, and the attempts themselves', async () => {++ const id = await makeTask();++ // The task ends, so the stream ends with it.++ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [id]);++ const job = { id: row.job_id, state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] }, error: null };++ await t.ctx.tasks.applyEvent(row, { state: 'succeeded', job });++ await t.ctx.tasks.afterEnd(id);++ const s = await call('GET', `/tasks/${id}/events`);++ assert.strictEqual(s.status, 200);++ assert.ok(s.text.includes('event: line'), s.text.slice(0, 300));++ const seqs = [...s.text.matchAll(/^id: (\d+)$/gm)].map((m) => Number(m[1]));++ assert.deepStrictEqual(seqs.slice(0, 3), [1, 2, 3], 'the lines in order, each with its seq');++ const bodies = [...s.text.matchAll(/^data: (.*)$/gm)].map((m) => JSON.parse(m[1]));++ const lines = bodies.filter((b) => b.seq);++ assert.deepStrictEqual(lines.slice(0, 3).map((l) => l.attempt_n), [1, 1, 2], 'each line says which attempt it belongs to');++ assert.deepStrictEqual(lines.slice(0, 3).map((l) => l.body.text), ['first', 'second', 'third']);++ const attempts = bodies.filter((b) => b.type);++ assert.strictEqual(attempts[0].type, 'attempt_started');++ assert.strictEqual(attempts[attempts.length - 1].type, 'attempt_ended');++ assert.strictEqual(attempts[attempts.length - 1].state, 'done');++ });++++ await check('a client that reconnects gets only what came after the line it last saw', async () => {++ const id = await makeTask();++ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [id]);++ const job = { id: row.job_id, state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] }, error: null };++ await t.ctx.tasks.applyEvent(row, { state: 'succeeded', job });++ await t.ctx.tasks.afterEnd(id);++ const s = await call('GET', `/tasks/${id}/events`, { headers: { 'Last-Event-ID': '2' } });++ assert.strictEqual(s.status, 200);++ const seqs = [...s.text.matchAll(/^id: (\d+)$/gm)].map((m) => Number(m[1]));++ assert.deepStrictEqual(seqs[0], 3, 'only the line after the last one it saw');++ assert.ok(!seqs.includes(1) && !seqs.includes(2), 'and not the ones before');++ assert.ok(!s.text.includes('"first"') && !s.text.includes('"second"'), 'and not the ones before');++ });++++ await check('the heartbeat keeps the connection alive', async () => {++ const id = await makeTask();++ const s = await stream(`/api/v1/tasks/${id}/events`, {++ headers: { Authorization: `Bearer ${actor}`, 'X-OV-Subject': ada.subject },++ until: ': heartbeat',++ ms: 5000,++ });++ assert.ok(s.text.includes(': heartbeat'), s.text.slice(0, 200));++ });++++ await check('the auth of the stream: no token, another person, another service', async () => {++ const id = await makeTask();++ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { token: null })).status, 401);++ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { subject: bo.subject })).status, 404);++ const other = t.network.mintService('svc:other', 'openvibe.codes', CAPS);++ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { token: other })).status, 404);++ // The owner needs no token (the task runs, so the stream is read until the heartbeat).++ const mine = await stream(`/api/v1/tasks/${id}/events`, { headers: { cookie: t.signIn(ada) }, until: ': heartbeat', ms: 5000 });++ assert.strictEqual(mine.status, 200);++ assert.ok(mine.text.includes('event: line'), mine.text.slice(0, 200));++ });++ } finally {++ await t.close();++ }++ done();++})().catch((e) => { console.error(e); process.exit(1); });@@@ -92,12 +92,24 @@ const { boot, check, done } = require('./helpers/boot');await db.exec('UPDATE run_tasks SET updated_at = updated_at - $2 WHERE id = $1', [task.id, 11 * 60_000]);const r = await call('GET', `/tasks/${task.id}`);assert.strictEqual(r.status, 200, JSON.stringify(r.json));-- const v = validate('codes.task@1', r.json);++ // The read is the task (its own fields at the top level, as codes.task@1 had them), with its++ // attempts and its outcome added to it: codes.task@2 (stage C2).++ const v = validate('codes.task@2', r.json);assert.ok(v.valid, JSON.stringify(v.errors));++ const v1doc = { ...r.json };++ delete v1doc.attempts; delete v1doc.outcome;++ const v1 = validate('codes.task@1', v1doc);++ assert.ok(v1.valid, 'the task itself still answers codes.task@1');assert.strictEqual(r.json.state, 'succeeded');++ assert.strictEqual(r.json.id, task.id, 'the task at the top level, not wrapped');assert.strictEqual(r.json.result.summary.length, 4000);assert.deepStrictEqual(r.json.result.files, [{ path: 'README', added: 1, removed: 0 }]);assert.deepStrictEqual(r.json.result.push, { pushed: false, reason: 'no token' });++ assert.strictEqual(r.json.attempts.length, 1, 'one attempt: the task as it was asked');++ assert.strictEqual(r.json.attempts[0].n, 1);++ assert.strictEqual(r.json.attempts[0].state, 'done');++ assert.ok(r.json.attempts[0].why, 'the attempt says why it happened');++ assert.strictEqual(r.json.outcome.state, 'done');});await check('another person, another service, or a task the person started by hand: 404', async () => {
@@ -20,9 +20,9 @@ One place to send a coding task to the agent that fits it — by what the task n
@@ -20,9 +20,9 @@ One place to send a coding task to the agent that fits it — by what the task n
| The harness | `openvibe-codes` ([bin/openvibe-codes.js](bin/openvibe-codes.js), [harness/](harness/)) | Runs on a person's own machine with their CLIs and keys; nothing goes through OpenVibe. See **Use the harness** below. |
| The harness | `openvibe-codes` ([bin/openvibe-codes.js](bin/openvibe-codes.js), [harness/](harness/)) | Runs on a person's own machine with their CLIs and keys; nothing goes through OpenVibe. See **Use the harness** below. |
| Home | `/` | What the harness is, the agents in the catalog, how a task finds its agent, Improve OpenVibe, and an honest table of what works today and what comes next. |
| Home | `/` | What the harness is, the agents in the catalog, how a task finds its agent, Improve OpenVibe, and an honest table of what works today and what comes next. |
| Get started | `/start` | Install, the agents and what each needs, the commands, permission levels, hand-offs, the event format and using it from code. |
| Get started | `/start` | Install, the agents and what each needs, the commands, permission levels, hand-offs, the event format and using it from code. |
-| Run on your machine | `/run`, `/run/tasks/:id` ([server/http/run.js](server/http/run.js)) | A signed-in person starts a coding task in any browser and their own machine runs it: an OpenVibe Run job (class `code`, artifact `[email protected]`, `trust: user-owned`) that Codes starts for them in their project (project delegation: `X-OV-Subject`, `X-OV-Project`; [server/domain/run-client.js](server/domain/run-client.js)). Run places it only on a machine of theirs (added on openvibe.bot/machines), whose OpenVibe Node runs `openvibe-codes job` ([harness/job.js](harness/job.js)) in its job sandbox: clone, run the agents with hand-offs, hand back the summary, the changed files and the patch, and, when asked and the machine holds `GITHUB_TOKEN`, push a new branch (never the one it started from) and open a pull request. The page streams the job's lines (Run's job stream through Codes) and keeps the result (`run_tasks`, migration 0003). Each task is its owner's; account export and deletion cover them (ADR-033, [server/account-data.js](server/account-data.js)). The task page shows the agents' **session** with openvibe-shared/agent-session.js: steps grouped, each step's command, change and output (live while it runs), helpers nested, hand-offs and the plan; after the end, the session from `codes_session_lines` (migration 0007): the lines are kept as the stream arrives, back-filled from the machine's final log, and replayed at `GET /run/tasks/:id/session.json?after=` (the owner, or a service with `codes.task.read` and `X-OV-Subject`); tasks from before keep the old `log`. Run's `run.job.*` events finalize a task at `POST /internal/events` ([server/events-consumer.js](server/events-consumer.js), no polling but a 10-minute fallback), and every state change is published as `codes.task.queued|running|succeeded|failed|cancelled` from an outbox ([server/events/outbox.js](server/events/outbox.js), contracts 0.232.0). |
+| Run on your machine | `/run`, `/run/tasks/:id` ([server/http/run.js](server/http/run.js)) | A signed-in person starts a coding task in any browser and their own machine runs it: an OpenVibe Run job (class `code`, artifact `[email protected]`, `trust: user-owned`) that Codes starts for them in their project (project delegation: `X-OV-Subject`, `X-OV-Project`; [server/domain/run-client.js](server/domain/run-client.js)). Run places it only on a machine of theirs (added on openvibe.bot/machines), whose OpenVibe Node runs `openvibe-codes job` ([harness/job.js](harness/job.js)) in its job sandbox: clone, run the agents with hand-offs, hand back the summary, the changed files and the patch, and, when asked and the machine holds `GITHUB_TOKEN`, push a new branch (never the one it started from) and open a pull request. The page streams the job's lines (Run's job stream through Codes) and keeps the result (`run_tasks`, migration 0003). Each task is its owner's; account export and deletion cover them (ADR-033, [server/account-data.js](server/account-data.js)). The task page shows the **timeline of attempts** above the agents' **session** (each attempt with its why: "Codex hit its limit, so Command Code took over with what it had done") and the session with openvibe-shared/agent-session.js: steps grouped, each step's command, change and output (live while it runs), helpers nested, hand-offs and the plan; after the end, the session from `codes_session_lines` (migration 0007): the lines are kept as the stream arrives, back-filled from the machine's final log, and replayed at `GET /run/tasks/:id/session.json?after=` (the owner, or a service with `codes.task.read` and `X-OV-Subject`); tasks from before keep the old `log`. Run's `run.job.*` events finalize a task at `POST /internal/events` ([server/events-consumer.js](server/events-consumer.js), no polling but a 10-minute fallback), and every state change is published as `codes.task.queued|running|succeeded|failed|cancelled` from an outbox ([server/events/outbox.js](server/events/outbox.js), contracts 0.232.0). |
| The agents | `/harnesses` | Every harness in [`server/data/harness-offers.json`](server/data/harness-offers.json) (validated at boot as `platform.harness-offer@1`, each model as `platform.agent-offer@1`) with its provider, address kind, what it can do, price and limits, and its models with context limits and token prices. No local path is ever shown. |
| The agents | `/harnesses` | Every harness in [`server/data/harness-offers.json`](server/data/harness-offers.json) (validated at boot as `platform.harness-offer@1`, each model as `platform.agent-offer@1`) with its provider, address kind, what it can do, price and limits, and its models with context limits and token prices. No local path is ever shown. |
-| Task API (services) | `POST /api/v1/tasks`, `GET /api/v1/tasks/:id`, `POST /api/v1/tasks/:id/cancel` ([server/http/tasks-api.js](server/http/tasks-api.js)) | Contracts 0.162 `codes.task.start` / `read` / `cancel`. A first-party service (OpenVibe.Actor) holding the capability starts a coding task for a person named in `X-OV-Subject`, with a Network service token for audience `openvibe.codes` (`CODES_AUDIENCE`). The task is the same one `/run` starts ([server/domain/run-tasks.js](server/domain/run-tasks.js)): it runs on that person's own machine, in the project of their latest `/run` task (none yet → 409 `codes.no_project`). It shows on their `/run` page. `idempotency_key` answers a retry with the first task (200, `created: false`). A service reads and cancels only the tasks it started for that person (otherwise 404). The result the machine returned is passed on only as `codes.task@1` allows, each field checked and cut to its limit (migration 0005: `via`, `idempotency_key`). Calls count against the person's budget: starts and cancels 6 a minute, 60 an hour. |
+| Task API (services) | `POST /api/v1/tasks`, `GET /api/v1/tasks/:id`, `GET /api/v1/tasks/:id/events`, `POST /api/v1/tasks/:id/continue`, `POST /api/v1/tasks/:id/stop`, `POST /api/v1/tasks/:id/cancel` ([server/http/tasks-api.js](server/http/tasks-api.js)) | Contracts 0.162 `codes.task.start` / `read` / `cancel`. A first-party service (OpenVibe.Actor) holding the capability starts a coding task for a person named in `X-OV-Subject`, with a Network service token for audience `openvibe.codes` (`CODES_AUDIENCE`). The task is the same one `/run` starts ([server/domain/run-tasks.js](server/domain/run-tasks.js)): it runs on that person's own machine, in the project of their latest `/run` task (none yet → 409 `codes.no_project`). It shows on their `/run` page. `idempotency_key` answers a retry with the first task (200, `created: false`). A service reads and cancels only the tasks it started for that person (otherwise 404); the owner needs no token for their own tasks. The result the machine returned is passed on only as `codes.task@1` allows, each field checked and cut to its limit (migration 0005: `via`, `idempotency_key`). Calls count against the person's budget: starts and cancels 6 a minute, 60 an hour. **A task's attempts** (stage C2, migration 0010 `codes_attempts`): every Run job is one attempt, so a continue, a switch to another agent and a wrap-up are attempts on one timeline, never new tasks. `GET /api/v1/tasks/:id` answers `codes.task@2`: the task's own fields at the top level, exactly as `codes.task@1` had them, and, added to them, `attempts[]` (the timeline) and `outcome` — `done`, `unfinished` (with what is left) or `stopped`. `GET /api/v1/tasks/:id/events` is the task's event stream (server-sent events): the session's lines in order with the attempt each belongs to, the attempts themselves (started, ended with why), a heartbeat every 15 s, and a resume with `Last-Event-ID`. `POST /api/v1/tasks/:id/continue` starts the next attempt now (with `agent` it switches), `POST /api/v1/tasks/:id/stop` stops the running attempt and nothing carries it on. |
| Agent API | `GET /api/v1/harnesses`, `POST /api/v1/harnesses/route` | Public, no key. The router takes `{ task, requirements? }` (`task` is `edit`, `review`, `long` or `resume`; `requirements.capabilities` adds hard constraints) and answers `{ selected, reasons, candidates }` from `openvibe-sdk/placement` — hard constraints first (a missing capability excludes an agent, and says so), then cost. RFC 9457 problems (`harness.invalid`, `harness.task_unknown`). |
| Agent API | `GET /api/v1/harnesses`, `POST /api/v1/harnesses/route` | Public, no key. The router takes `{ task, requirements? }` (`task` is `edit`, `review`, `long` or `resume`; `requirements.capabilities` adds hard constraints) and answers `{ selected, reasons, candidates }` from `openvibe-sdk/placement` — hard constraints first (a missing capability excludes an agent, and says so), then cost. RFC 9457 problems (`harness.invalid`, `harness.task_unknown`). |
| Improve OpenVibe | `/improve` | Every OpenVibe repository, from the pinned contracts' service manifests (the services that run, then the shared libraries), and the path from a change to a reviewed pull request. |
| Improve OpenVibe | `/improve` | Every OpenVibe repository, from the pinned contracts' service manifests (the services that run, then the shared libraries), and the path from a change to a reviewed pull request. |
| Community | `/policy`, `/policy/<document>` | The code of conduct ([CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md)), the contribution guide ([CONTRIBUTING.md](CONTRIBUTING.md)), the [contributor ladder](docs/governance/contributor-ladder.md) and the [moderation policy](docs/governance/moderation.md), rendered from those files. All four are **drafts pending owner review**: marked at the top, shown with a notice, not indexed; deleting the draft line publishes one. They govern the repositories and the apps and releases on OpenVibe.Services alike. |
| Community | `/policy`, `/policy/<document>` | The code of conduct ([CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md)), the contribution guide ([CONTRIBUTING.md](CONTRIBUTING.md)), the [contributor ladder](docs/governance/contributor-ladder.md) and the [moderation policy](docs/governance/moderation.md), rendered from those files. All four are **drafts pending owner review**: marked at the top, shown with a notice, not indexed; deleting the draft line publishes one. They govern the repositories and the apps and releases on OpenVibe.Services alike. |
@@ -84,7 +84,7 @@ openvibe-codes resume <run id> "now add a test"
@@ -84,7 +84,7 @@ openvibe-codes resume <run id> "now add a test"
- **OpenVibe.Network** — SSO (OAuth client `codes`, PKCE S256) and its JWKS.
- **OpenVibe.Network** — SSO (OAuth client `codes`, PKCE S256) and its JWKS.
- **PostgreSQL 18 and Valkey 9** (OpenVibe.Host `roles/data/`, ADR-035): `openvibe-sdk/db`; Valkey holds the per-actor limit counters (optional).
- **PostgreSQL 18 and Valkey 9** (OpenVibe.Host `roles/data/`, ADR-035): `openvibe-sdk/db`; Valkey holds the per-actor limit counters (optional).
-- **openvibe-contracts v0.143.0**, **openvibe-sdk v0.67.0**, **openvibe-shared v4.23.0** (pinned tag tarballs).
+- **openvibe-contracts v0.243.0**, **openvibe-sdk v0.73.0**, **openvibe-shared v4.48.0** (pinned tag tarballs).
No path in Codes sends or accepts a shared loopback key (tested by grep and at runtime).
No path in Codes sends or accepts a shared loopback key (tested by grep and at runtime).
@@ -158,7 +158,7 @@ Reporting a vulnerability: [SECURITY.md](SECURITY.md).
@@ -158,7 +158,7 @@ Reporting a vulnerability: [SECURITY.md](SECURITY.md).
Part of the [OpenVibe network](https://openvibe.network). Built in the open by [OpenVibers](https://github.com/OpenVibers).
Part of the [OpenVibe network](https://openvibe.network). Built in the open by [OpenVibers](https://github.com/OpenVibers).
<!-- versions:start -->
<!-- versions:start -->
-- openvibe-contracts: v0.232.0
+- openvibe-contracts: v0.243.0
- openvibe-sdk: v0.73.0
- openvibe-sdk: v0.73.0
- openvibe-shared: v4.48.0
- openvibe-shared: v4.48.0
<!-- versions:end -->
<!-- versions:end -->
@@ -7,11 +7,11 @@
@@ -7,11 +7,11 @@
"runtime": "node 22, express 4, PostgreSQL 18 (openvibe-sdk/db, async) + Valkey 9",
"runtime": "node 22, express 4, PostgreSQL 18 (openvibe-sdk/db, async) + Valkey 9",
"port": 4900,
"port": 4900,
"domain": "openvibe.codes",
"domain": "openvibe.codes",
- "contracts": "openvibe-contracts v0.232.0",
+ "contracts": "openvibe-contracts v0.243.0",
"sdk": "openvibe-sdk v0.73.0",
"sdk": "openvibe-sdk v0.73.0",
"shared": "openvibe-shared v4.48.0",
"shared": "openvibe-shared v4.48.0",
"packages": [
"packages": [
- "openvibe-contracts v0.232.0",
+ "openvibe-contracts v0.243.0",
"openvibe-sdk v0.73.0",
"openvibe-sdk v0.73.0",
"openvibe-shared v4.48.0"
"openvibe-shared v4.48.0"
],
],
@@ -84,8 +84,13 @@ function parseArgs(raw, env = process.env) {
@@ -84,8 +84,13 @@ function parseArgs(raw, env = process.env) {
} : null;
} : null;
if (push && push.branch && (!BRANCH.test(push.branch) || push.branch.startsWith('-'))) throw new Error('push.branch must be a branch name');
if (push && push.branch && (!BRANCH.test(push.branch) || push.branch.startsWith('-'))) throw new Error('push.branch must be a branch name');
if (push && push.branch && PROTECTED.test(push.branch)) throw new Error(`push.branch must be a new branch of the task's own, never ${push.branch}`);
if (push && push.branch && PROTECTED.test(push.branch)) throw new Error(`push.branch must be a new branch of the task's own, never ${push.branch}`);
+ // The session a resume continues: the job's own input, or RESUME in the environment (a task carried on in the
+ // same conversation). It is a harness's session id, never a path and never a credential.
+ const resumeIn = a.resume == null || a.resume === '' ? env.RESUME : a.resume;
+ const resume = typeof resumeIn === 'string' && /^[A-Za-z0-9._:-]{1,200}$/.test(resumeIn) ? resumeIn : null;
return {
return {
repo: repo.toString(), ref, prompt, permission, push,
repo: repo.toString(), ref, prompt, permission, push,
+ ...(resume ? { resume } : {}),
agent: typeof a.agent === 'string' && /^[a-z0-9-]{1,40}$/.test(a.agent) ? a.agent : null,
agent: typeof a.agent === 'string' && /^[a-z0-9-]{1,40}$/.test(a.agent) ? a.agent : null,
model: typeof a.model === 'string' && a.model.length <= 100 ? a.model : null,
model: typeof a.model === 'string' && a.model.length <= 100 ? a.model : null,
};
};
@@ -129,7 +134,7 @@ const resultText = (c) => (typeof c === 'string' ? c : Array.isArray(c) ? c.map(
@@ -129,7 +134,7 @@ const resultText = (c) => (typeof c === 'string' ? c : Array.isArray(c) ? c.map(
function compact(e) {
function compact(e) {
const cut = (s) => cutTo(s, MAX_TEXT);
const cut = (s) => cutTo(s, MAX_TEXT);
const parent = e.parent_tool_use_id ? { parent: String(e.parent_tool_use_id) } : {};
const parent = e.parent_tool_use_id ? { parent: String(e.parent_tool_use_id) } : {};
- if (e.type === 'system' && e.subtype === 'init') return { t: 'agent', agent: e.agent, model: e.model || null };
+ if (e.type === 'system' && e.subtype === 'init') return { t: 'agent', agent: e.agent, model: e.model || null, session: e.session_id || null };
if (e.type === 'system' && e.subtype === 'handoff') return { t: 'handoff', from: e.from, ...(e.to ? { to: e.to } : {}), reason: cut(e.reason) };
if (e.type === 'system' && e.subtype === 'handoff') return { t: 'handoff', from: e.from, ...(e.to ? { to: e.to } : {}), reason: cut(e.reason) };
if (e.type === 'system' && (e.subtype === 'error' || e.subtype === 'warning')) return { t: 'note', level: e.subtype, text: cut(e.text) };
if (e.type === 'system' && (e.subtype === 'error' || e.subtype === 'warning')) return { t: 'note', level: e.subtype, text: cut(e.text) };
if (e.type === 'system' && e.subtype === 'tool_progress') return { t: 'tool_progress', id: String(e.tool_use_id || ''), chunk: clean(cutTo(e.text, MAX_OUTPUT)) };
if (e.type === 'system' && e.subtype === 'tool_progress') return { t: 'tool_progress', id: String(e.tool_use_id || ''), chunk: clean(cutTo(e.text, MAX_OUTPUT)) };
@@ -266,6 +271,9 @@ async function runJob({ input, harness, emit, writeResult, env = process.env, tm
@@ -266,6 +271,9 @@ async function runJob({ input, harness, emit, writeResult, env = process.env, tm
const run = harness.run({
const run = harness.run({
prompt: job.prompt, task: 'edit', agent: job.agent, model: job.model, cwd: work,
prompt: job.prompt, task: 'edit', agent: job.agent, model: job.model, cwd: work,
permission: job.permission, handoff: !job.agent, maxAttempts: 3,
permission: job.permission, handoff: !job.agent, maxAttempts: 3,
+ // A task carried on in the same conversation resumes its session (RESUME in the environment, or the job's
+ // own resume field).
+ resume: job.resume || null,
});
});
let last = null;
let last = null;
for await (const e of run) {
for await (const e of run) {
new file mode 100644
new file mode 100644
@@ -0,0 +1,109 @@
@@ -0,0 +1,109 @@
+-- phase: expand
+-- Stage C2 (openvibe.codes v1): a task's attempts. A continue, a switch to another agent and a wrap-up are attempts on
+-- one task's timeline, never new tasks with cryptic names (the portal's "-c1"/"-r2" confusion of 10-30). Every Run job
+-- is one attempt: the attempt carries the job, the agent, the machine and the harness session a resume needs.
+--
+-- n the attempt's place on the timeline (1, 2, …)
+-- state queued | running | done | failed | stopped | budget-stopped | interrupted
+-- reason limit | auth | capacity | refused | error | budget | stopped | hung | context (why it ended)
+-- why one sentence in plain words: why this attempt happened (for the timeline)
+-- kind first | continue | switch | wrap-up | manual: how this attempt came to be
+-- follow_up_of the attempt whose end started this one. It is unique, so one ending can start only one follow-up:
+-- the fence against the portal's 10-30 bug, where two racing ticks carried one job on twice.
+--
+-- run_tasks grows the task's own verdict: outcome ('done' | 'unfinished' | 'stopped', NULL while it runs), outcome_note
+-- (what is left, in plain words) and attempts_count. codes_session_lines grows attempt_n, so every line of the session
+-- says which attempt it belongs to.
+--
+-- Every task that exists now becomes attempt 1 of itself: the timeline of an old task is not empty.
+
+CREATE TABLE IF NOT EXISTS codes_attempts (
+ id text COLLATE "C" PRIMARY KEY, -- att_<ULID>
+ task_id text COLLATE "C" NOT NULL, -- tsk_<ULID>
+ n integer NOT NULL, -- 1, 2, … the attempt's place on the timeline
+ agent text COLLATE "C", -- the harness (claude-code, codex, …) or 'auto'
+ model text COLLATE "C",
+ machine text COLLATE "C", -- the node's id, or 'sandbox'
+ job_id text COLLATE "C", -- Run's job_<ULID> for this attempt
+ session text COLLATE "C", -- the harness session id, for a resume
+ state text COLLATE "C" NOT NULL, -- queued, running, done, failed, stopped, budget-stopped, interrupted
+ reason text COLLATE "C", -- limit, auth, capacity, refused, error, budget, stopped, hung, context
+ why text COLLATE "C", -- one sentence in plain words
+ kind text COLLATE "C" NOT NULL DEFAULT 'first',-- first | continue | switch | wrap-up | manual
+ cost_usd numeric(14,6),
+ tokens bigint,
+ steps integer,
+ started_at timestamptz,
+ finished_at timestamptz,
+ created_at timestamptz NOT NULL DEFAULT now(),
+ follow_up_of text COLLATE "C" -- the attempt whose end started this one
+);
+CREATE UNIQUE INDEX IF NOT EXISTS codes_attempts_task_n ON codes_attempts (task_id, n);
+-- One ending, one follow-up: the unique index is the fence (the second insert is a no-op).
+CREATE UNIQUE INDEX IF NOT EXISTS codes_attempts_follow_up ON codes_attempts (follow_up_of) WHERE follow_up_of IS NOT NULL;
+-- A Run job is one attempt: the job of an attempt is found by it.
+CREATE INDEX IF NOT EXISTS codes_attempts_job ON codes_attempts (job_id) WHERE job_id IS NOT NULL;
+CREATE INDEX IF NOT EXISTS codes_attempts_task ON codes_attempts (task_id, n);
+
+ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS outcome text COLLATE "C"; -- done | unfinished | stopped
+ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS outcome_note text COLLATE "C"; -- what is left, in plain words
+ALTER TABLE run_tasks ADD COLUMN IF NOT EXISTS attempts_count integer NOT NULL DEFAULT 0;
+ALTER TABLE codes_session_lines ADD COLUMN IF NOT EXISTS attempt_n integer;
+
+-- Every task that exists becomes attempt 1 of itself. The id is the task's own id hashed into the ULID alphabet
+-- (26 characters of Crockford base32), so the same task always gets the same attempt and a rerun of the migration
+-- changes nothing.
+INSERT INTO codes_attempts (id, task_id, n, agent, job_id, state, reason, why, kind, started_at, finished_at, created_at)
+SELECT 'att_' || upper(substr(md5(t.id), 1, 26)),
+ t.id,
+ 1,
+ t.agent,
+ t.job_id,
+ CASE t.state
+ WHEN 'queued' THEN 'queued'
+ WHEN 'placed' THEN 'queued'
+ WHEN 'running' THEN 'running'
+ WHEN 'succeeded' THEN 'done'
+ WHEN 'cancelled' THEN 'stopped'
+ ELSE 'failed'
+ END,
+ CASE t.state
+ WHEN 'cancelled' THEN 'stopped'
+ WHEN 'expired' THEN 'capacity'
+ WHEN 'failed' THEN 'error'
+ ELSE NULL
+ END,
+ CASE t.state
+ WHEN 'succeeded' THEN 'the first attempt of this task finished the work'
+ WHEN 'cancelled' THEN 'the first attempt of this task was stopped on purpose'
+ WHEN 'expired' THEN 'no machine took the first attempt in time'
+ WHEN 'failed' THEN 'the first attempt of this task failed'
+ WHEN 'running' THEN 'the first attempt of this task is running'
+ ELSE 'the first attempt of this task is waiting for a machine'
+ END,
+ 'first',
+ t.started_at,
+ t.finished_at,
+ to_timestamp(GREATEST(t.created_at, 0) / 1000.0)
+FROM run_tasks t
+WHERE NOT EXISTS (SELECT 1 FROM codes_attempts a WHERE a.task_id = t.id);
+
+UPDATE run_tasks t
+SET outcome = CASE t.state
+ WHEN 'succeeded' THEN 'done'
+ WHEN 'cancelled' THEN 'stopped'
+ WHEN 'failed' THEN 'unfinished'
+ WHEN 'expired' THEN 'unfinished'
+ ELSE NULL
+ END,
+ outcome_note = CASE t.state
+ WHEN 'succeeded' THEN 'the task finished before attempts were counted'
+ WHEN 'cancelled' THEN 'the task was stopped on purpose'
+ WHEN 'failed' THEN 'the attempt failed; what it left is in the session'
+ WHEN 'expired' THEN 'no machine took the task in time'
+ ELSE NULL
+ END,
+ attempts_count = 1
+WHERE t.attempts_count = 0;
+
+UPDATE codes_session_lines SET attempt_n = 1 WHERE attempt_n IS NULL;
@@ -15,7 +15,7 @@
@@ -15,7 +15,7 @@
"express-rate-limit": "^7.5.0",
"express-rate-limit": "^7.5.0",
"helmet": "^8.1.0",
"helmet": "^8.1.0",
"iovalkey": "^0.4.0",
"iovalkey": "^0.4.0",
- "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",
+ "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0",
"openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0",
"openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0",
"openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0",
"openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0",
"pg": "^8.23.0"
"pg": "^8.23.0"
@@ -1149,9 +1149,9 @@
@@ -1149,9 +1149,9 @@
}
}
},
},
"node_modules/openvibe-contracts": {
"node_modules/openvibe-contracts": {
- "version": "0.232.0",
- "resolved": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",
- "integrity": "sha512-6HI5nn//nAUEDdG4xRTAlcb40oc/6yO1nx7R5exGZ6Bo3mL5fJ5szw9IONddoLN6qJsO1CXzB5TzTqkBAh3OzQ==",
+ "version": "0.243.0",
+ "resolved": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0",
+ "integrity": "sha512-ZFrwTpgALUEOIvdYi2r654murdEwludenAZ5u3CuvPMVcYuu/3kC+J20jVeqPaCFKYkzRzLyO8fH/iOdh44aYg==",
"license": "AGPL-3.0-only",
"license": "AGPL-3.0-only",
"dependencies": {
"dependencies": {
"ajv": "^8.17.1",
"ajv": "^8.17.1",
@@ -30,7 +30,7 @@
@@ -30,7 +30,7 @@
"express-rate-limit": "^7.5.0",
"express-rate-limit": "^7.5.0",
"helmet": "^8.1.0",
"helmet": "^8.1.0",
"iovalkey": "^0.4.0",
"iovalkey": "^0.4.0",
- "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.232.0",
+ "openvibe-contracts": "https://codeload.github.com/OpenVibers/OpenVibe.Contracts/tar.gz/refs/tags/v0.243.0",
"openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0",
"openvibe-sdk": "https://codeload.github.com/OpenVibers/OpenVibe.SDK/tar.gz/refs/tags/v0.73.0",
"openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0",
"openvibe-shared": "https://codeload.github.com/OpenVibers/OpenVibe.Shared/tar.gz/refs/tags/v4.48.0",
"pg": "^8.23.0"
"pg": "^8.23.0"
@@ -32,6 +32,7 @@ const { createTasksApi } = require('./http/tasks-api');
@@ -32,6 +32,7 @@ const { createTasksApi } = require('./http/tasks-api');
const { createActionsApi } = require('./http/actions-api');
const { createActionsApi } = require('./http/actions-api');
const { createRunClient } = require('./domain/run-client');
const { createRunClient } = require('./domain/run-client');
const { createRunTasks } = require('./domain/run-tasks');
const { createRunTasks } = require('./domain/run-tasks');
+const { createAttempts } = require('./domain/attempts');
const { createSessionLines } = require('./domain/session-lines');
const { createSessionLines } = require('./domain/session-lines');
const { createOutbox, createPublisher } = require('./events/outbox');
const { createOutbox, createPublisher } = require('./events/outbox');
const { createEventsConsumer } = require('./events-consumer');
const { createEventsConsumer } = require('./events-consumer');
@@ -66,8 +67,10 @@ async function createApp(opts = {}) {
@@ -66,8 +67,10 @@ async function createApp(opts = {}) {
const sessions = createSessionLines({ store, config, log });
const sessions = createSessionLines({ store, config, log });
const outbox = createOutbox({ store, log });
const outbox = createOutbox({ store, log });
const run = opts.run || createRunClient({ config, fetchImpl: fetchImpl || globalThis.fetch });
const run = opts.run || createRunClient({ config, fetchImpl: fetchImpl || globalThis.fetch });
- const tasks = createRunTasks({ store, run, config, log, sessions, outbox });
- const ctx = { config, store, keys, sso, harnesses, log, sessions, outbox, run, tasks };
+ // The task's attempts (stage C2): every Run job is one attempt, and the engine decides what happens next.
+ const attempts = createAttempts({ store, config, log, harnesses });
+ const tasks = createRunTasks({ store, run, config, log, sessions, outbox, attempts });
+ const ctx = { config, store, keys, sso, harnesses, log, sessions, outbox, run, tasks, attempts };
ctx.forgejo = createForgejoClient({ config, log });
ctx.forgejo = createForgejoClient({ config, log });
const app = express();
const app = express();
@@ -81,6 +81,18 @@ function load(env = process.env) {
@@ -81,6 +81,18 @@ function load(env = process.env) {
// OpenVibe.Run (plan T16): a coding task on the person's own machine is a Run job Codes starts for them.
// OpenVibe.Run (plan T16): a coding task on the person's own machine is a Run job Codes starts for them.
run: { url: trim(env.CODES_RUN_URL || 'http://127.0.0.1:4920') },
run: { url: trim(env.CODES_RUN_URL || 'http://127.0.0.1:4920') },
+ // A task's own budget (engine/budget.js): steps (tool calls), minutes (wall clock) and dollars. A limit that is
+ // not set is no limit — 0 means "no limit", never a zero budget. Steps and minutes buy one wrap-up turn.
+ budget: {
+ steps: int(env.CODES_TASK_STEPS, 0),
+ minutes: int(env.CODES_TASK_MINUTES, 0),
+ usd: Number(env.CODES_TASK_USD || 0) || 0,
+ },
+ // The task's timeline (stage C2): how many automatic continues one task may have before it needs a person.
+ task: { maxContinues: Math.max(0, int(env.CODES_TASK_MAX_CONTINUES, 2)) },
+ // The event stream of a task (GET /api/v1/tasks/:id/events): the heartbeat that keeps a connection alive.
+ sse: { heartbeatMs: Math.max(1000, int(env.CODES_SSE_HEARTBEAT_MS, 15_000)) },
+
};
};
}
}
new file mode 100644
new file mode 100644
@@ -0,0 +1,418 @@
@@ -0,0 +1,418 @@
+'use strict';
+
+/**
+ * A coding task's attempts (codes_attempts): the timeline of one task.
+ *
+ * A continue, a switch to another agent and a wrap-up are attempts on one task's timeline, never new tasks with
+ * cryptic names (the portal's "-c1"/"-r2" confusion of 10-30). Every Run job is one attempt: the attempt carries the
+ * job, the agent, the machine and the harness session a resume needs.
+ *
+ * The engine decides what happens next (engine/chain.js outcome/nextStep, engine/budget.js wrapUp): this module turns
+ * its decision into a plan — a follow-up attempt, or the task's outcome in plain words. It never starts a Run job
+ * itself (server/domain/run-tasks.js does that, outside a transaction) and it never decides twice for one ending:
+ * the follow-up's follow_up_of is unique, so one ending can start only one follow-up.
+ *
+ * createAttempts({ store, config, log, harnesses })
+ * list(taskId) the attempts of a task, in order
+ * get(id) one attempt
+ * create({ task, …, followUpOf }) one attempt; the same followUpOf answers the same attempt (the fence)
+ * bind(id, { job_id, state }) the Run job of an attempt
+ * end(id, { state, reason, … }) an attempt ended: its state, its reason, its cost
+ * decide({ task, attempts, ending, result }) the engine's decision, as a plan
+ * present(attempt) the attempt as the API shows it
+ */
+
+const { chain, budget, router } = require('../../engine');
+
+/** Codes' harness ids and the engine's short runner ids for the same harness. */
+const RUNNER_OF = { 'claude-code': 'claude', codex: 'cx', 'command-code': 'cc', opencode: 'oc', deepseek: 'dsh' };
+const HARNESS_OF = { claude: 'claude-code', cx: 'codex', cc: 'command-code', oc: 'opencode', dsh: 'deepseek' };
+
+/** The order a runner switch tries the harnesses in (the ladder's order, flattened). */
+const FALLBACK_ORDER = ['cx', 'claude', 'cc', 'oc', 'dsh'];
+
+/** Which harnesses may take which kind of task, cheapest first (the engine's ladder, Codes' harnesses). */
+const TIER_ORDER = {
+ basic: ['opencode', 'command-code', 'deepseek'],
+ research: ['codex', 'opencode', 'deepseek'],
+ standard: ['codex', 'claude-code', 'command-code', 'opencode', 'deepseek'],
+ hard: ['codex', 'claude-code', 'deepseek'],
+ design: ['codex', 'claude-code', 'opencode'],
+};
+const TIER_EFFORT = { basic: 'low', research: 'medium', standard: 'high', hard: 'high', design: 'high' };
+
+/** The attempt's own states, and the reasons it may have ended. */
+const STATES = ['queued', 'running', 'done', 'failed', 'stopped', 'budget-stopped', 'interrupted'];
+const REASONS = ['limit', 'auth', 'capacity', 'refused', 'error', 'budget', 'stopped', 'hung', 'context'];
+const KINDS = ['first', 'continue', 'switch', 'wrap-up', 'restart', 'manual'];
+
+/** The engine's own words for a reason: the timeline reads like a person wrote it. */
+const REASON_TEXT = chain.REASON_TEXT;
+
+/**
+ * The text the engine classifies a reason from. The engine reads the reason of a failed attempt out of what it said
+ * (engine/chain.js reason), so the text has to say the same thing our reason column says.
+ */
+const ERROR_TEXT = {
+ limit: 'usage limit reached',
+ auth: 'authentication failed: log in again',
+ capacity: 'the model was at capacity',
+ refused: 'turned off in the agent portal',
+ error: 'the work failed',
+ budget: 'stopped at its limit',
+ stopped: 'stopped on purpose',
+ context: 'the context window is too big',
+ hung: 'it went quiet for too long',
+};
+
+/** The engine's status for our state (the two vocabularies are the same, except 'queued' and 'running'). */
+const ENGINE_STATE = {
+ queued: 'queued', running: 'running', done: 'done', failed: 'failed',
+ stopped: 'stopped', 'budget-stopped': 'budget-stopped', interrupted: 'interrupted',
+};
+
+/** The attempt's state for a task's state (a task from before the attempts still has one). */
+const TASK_STATE = {
+ queued: 'queued', placed: 'queued', running: 'running', succeeded: 'done',
+ failed: 'failed', cancelled: 'stopped', expired: 'failed',
+};
+
+const iso = (v) => {
+ if (!v) return null;
+ const t = v instanceof Date ? v.getTime() : Date.parse(String(v));
+ return Number.isFinite(t) ? new Date(t).toISOString() : null;
+};
+const clip = (s, n) => String(s == null ? '' : s).slice(0, n);
+const runnerOf = (agent) => RUNNER_OF[agent] || (agent || 'auto');
+const wordsOf = (reason) => REASON_TEXT[reason] || REASON_TEXT.error;
+
+function createAttempts({ store, config = {}, log = console, harnesses = null }) {
+ const db = store.db;
+ const catalog = harnesses || require('./harnesses').createHarnesses();
+ // How many automatic continues one task may have before it needs a person (CODES_TASK_MAX_CONTINUES, two).
+ const wanted = config.task ? config.task.maxContinues : undefined;
+ const maxContinues = Number.isFinite(Number(wanted)) ? Math.max(0, Number(wanted)) : 2;
+
+ const nameOf = (agent) => {
+ const offer = agent ? catalog.get(agent) : null;
+ return offer ? offer.name : (agent || 'the agent');
+ };
+
+ /** The attempts of a task, in order. A task written before the attempts (or by another path) still has a first one. */
+ async function list(taskId) {
+ if (!taskId) return [];
+ const rows = await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);
+ if (rows.length) return rows;
+ const task = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [taskId]);
+ if (!task) return [];
+ await db.exec(
+ `INSERT INTO codes_attempts (id, task_id, n, agent, job_id, state, why, kind, started_at, finished_at, created_at)
+ VALUES ($1, $2, 1, $3, $4, $5, $6, 'first', $7::timestamptz, $8::timestamptz, $9::timestamptz)
+ ON CONFLICT DO NOTHING`,
+ [store.newId('att'), task.id, task.agent || null, task.job_id || null, TASK_STATE[task.state] || 'failed',
+ 'the first attempt of this task', task.started_at || null, task.finished_at || null, store.iso()],
+ );
+ return await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);
+ }
+
+ /** One attempt. */
+ async function get(id) {
+ if (!id) return null;
+ return await db.maybe('SELECT * FROM codes_attempts WHERE id = $1', [id]);
+ }
+
+ /** The attempt a job belongs to, or null. */
+ async function byJob(jobId) {
+ if (!jobId) return null;
+ return await db.maybe('SELECT * FROM codes_attempts WHERE job_id = $1', [jobId]);
+ }
+
+ /**
+ * One attempt. The fence: follow_up_of is unique, so an ending that starts a follow-up twice answers the same
+ * attempt the second time and starts no second job. created tells the caller which of the two it was.
+ */
+ async function create({ task, agent = null, model = null, machine = null, session = null, why = null, kind = 'first', followUpOf = null, state = 'queued' }) {
+ if (!task || !task.id) throw new Error('an attempt belongs to a task');
+ for (let tries = 0; tries < 5; tries++) {
+ const n = await db.one('SELECT COALESCE(MAX(n), 0) + 1 AS n FROM codes_attempts WHERE task_id = $1', [task.id]);
+ const id = store.newId('att');
+ const inserted = await db.maybe(
+ `INSERT INTO codes_attempts (id, task_id, n, agent, model, machine, session, state, why, kind, follow_up_of, created_at)
+ VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12::timestamptz)
+ ON CONFLICT DO NOTHING RETURNING *`,
+ [id, task.id, n.n, agent, model, machine, session, state, why, kind, followUpOf, store.iso()],
+ );
+ if (inserted) {
+ await db.exec('UPDATE run_tasks SET attempts_count = $2 WHERE id = $1', [task.id, n.n]);
+ return { attempt: inserted, created: true };
+ }
+ // Someone else wrote it: the same follow-up answers the same attempt.
+ const seen = followUpOf ? await db.maybe('SELECT * FROM codes_attempts WHERE follow_up_of = $1', [followUpOf]) : null;
+ if (seen) return { attempt: seen, created: false };
+ }
+ throw new Error('the attempt could not be written');
+ }
+
+ /** The Run job of an attempt. */
+ async function bind(id, { job_id = null, state = null, started_at = null, machine = null, model = null } = {}) {
+ return await db.one(
+ `UPDATE codes_attempts SET job_id = COALESCE($2, job_id),
+ state = COALESCE($3, state),
+ machine = COALESCE($4, machine),
+ model = COALESCE($5, model),
+ started_at = COALESCE($6::timestamptz, started_at)
+ WHERE id = $1 RETURNING *`,
+ [id, job_id, state, machine, model, started_at],
+ );
+ }
+
+ /** An attempt ended: its state, why, what it cost. */
+ async function end(id, { state = 'failed', reason = null, why = null, cost_usd = null, tokens = null, steps = null, session = null, model = null, machine = null, finished_at = null } = {}) {
+ return await db.one(
+ `UPDATE codes_attempts SET state = $2,
+ reason = COALESCE($3, reason),
+ why = COALESCE($4, why),
+ cost_usd = COALESCE($5::numeric, cost_usd),
+ tokens = COALESCE($6, tokens),
+ steps = COALESCE($7, steps),
+ session = COALESCE($8, session),
+ model = COALESCE($9, model),
+ machine = COALESCE($10, machine),
+ finished_at = COALESCE($11::timestamptz, finished_at)
+ WHERE id = $1 RETURNING *`,
+ [id, state, reason, why, cost_usd == null ? null : String(cost_usd), tokens, steps, session, model, machine, finished_at || store.iso()],
+ );
+ }
+
+ /** The engine's own vocabulary for Codes' harnesses: the ladder, the health and the catalogs it routes with. */
+ function routing() {
+ const subagents = {}; const models = {}; const runners = [];
+ for (const [harnessId, runner] of Object.entries(RUNNER_OF)) {
+ const offer = catalog.get(harnessId);
+ if (!offer) continue;
+ subagents[runner] = { enabled: true, max: 4, model: null };
+ models[runner] = (offer.agents || []).map((a) => a.model).filter(Boolean);
+ runners.push(runner);
+ }
+ const health = {};
+ for (const r of runners) health[r] = { state: 'ok' };
+ const ladder = {};
+ for (const tier of router.TIERS) {
+ ladder[tier] = (TIER_ORDER[tier] || []).map((id) => {
+ const runner = RUNNER_OF[id];
+ const offer = catalog.get(id);
+ const agent = offer && offer.agents && offer.agents[0];
+ return [runner, agent ? agent.model : null, TIER_EFFORT[tier]];
+ }).filter(([runner, model]) => subagents[runner] && (!model || models[runner].includes(model)));
+ }
+ return {
+ cfg: { subagents, budget: config.budget || {}, fallback: { order: FALLBACK_ORDER.filter((r) => subagents[r]) } },
+ health, counts: {}, models, runners, ladder,
+ };
+ }
+
+ /**
+ * The next agent for a task, from the engine's router (engine/router.js): the tier of the brief decides the
+ * ladder, and the first harness of it that may take the task takes it. The agent that just stopped is skipped.
+ */
+ function nextAgent(prompt, ctx, exclude) {
+ const usable = (runner) => {
+ const sub = ctx.cfg.subagents[runner];
+ if (!sub || !sub.enabled) return false;
+ if ((ctx.counts[runner] || 0) >= sub.max) return false;
+ const h = ctx.health[runner];
+ return !!h && h.state === 'ok';
+ };
+ const picked = router.pick(prompt, ctx);
+ const tier = picked ? picked.tier : router.classify(prompt).tier;
+ const ladder = ctx.ladder[tier] || [];
+ const shape = (runner, model, effort) => ({
+ runner,
+ agent: HARNESS_OF[runner] || null,
+ model: model || null,
+ effort: router.fitEffort(runner, effort),
+ tier,
+ why: picked ? picked.why : `${tier} → ${runner}`,
+ });
+ if (picked && picked.runner !== exclude && usable(picked.runner)) return shape(picked.runner, picked.model, picked.effort);
+ for (const [runner, model, effort] of ladder) {
+ if (runner === exclude || !usable(runner)) continue;
+ const m = model && ctx.models[runner] && !ctx.models[runner].includes(model) ? null : model;
+ return shape(runner, m, effort);
+ }
+ return null;
+ }
+
+ /** One attempt as the engine sees it. */
+ function engineAttempt(a, all, facts = null) {
+ const n = Number(a.n) || 1;
+ const reason = a.reason && REASONS.includes(a.reason) ? a.reason : null;
+ const next = all.find((x) => x.follow_up_of && x.follow_up_of === a.id);
+ const prev = all.find((x) => Number(x.n) === n - 1);
+ const created = a.created_at ? Date.parse(String(a.created_at)) : 0;
+ return {
+ key: a.id,
+ runner: runnerOf(a.agent),
+ name: n === 1 ? 'a' : `a-r${n}`,
+ runnerName: nameOf(a.agent),
+ status: ENGINE_STATE[a.state] || 'failed',
+ // The order the engine reads the chain in: the attempt's place, never a tie.
+ updated: created + n,
+ attempt: n,
+ error: reason ? ERROR_TEXT[reason] : null,
+ // The engine reads the reason of a failed attempt out of what it said; Codes' attempts carry their own
+ // reason, and the engine only reads the runners it knows, so it is passed on.
+ reason: reason || null,
+ next: next ? next.id : null,
+ prev: prev ? prev.id : null,
+ parent: a.follow_up_of || null,
+ stopped: a.state === 'stopped',
+ budget: reason === 'budget' ? { hit: 'steps' } : undefined,
+ wrapUp: a.kind === 'wrap-up',
+ hung: reason === 'hung',
+ facts: {
+ session: a.session || null,
+ apiError: (facts && facts.apiError) || (reason ? ERROR_TEXT[reason] : null),
+ hasFinal: !!(facts && facts.hasFinal),
+ finalText: (facts && facts.finalText) || '',
+ },
+ };
+ }
+
+ /**
+ * What the job's log says about the attempt: its last words (the engine reads them to know whether the agent
+ * said it was done) and the error it ended with.
+ */
+ function factsOf(result) {
+ const d = result && typeof result === 'object' && !Array.isArray(result) ? result : {};
+ const log = Array.isArray(d.log) ? d.log : [];
+ let hasFinal = false; let finalText = ''; let apiError = '';
+ for (const line of log) {
+ if (!line || typeof line !== 'object') continue;
+ if (line.t === 'done') {
+ hasFinal = true;
+ if (line.text) finalText = String(line.text);
+ if (line.ok === false && !apiError) apiError = finalText || 'the agent ended without an answer';
+ } else if (line.t === 'note' && line.level === 'error' && !apiError) {
+ apiError = String(line.text || '');
+ }
+ }
+ return { hasFinal, finalText, apiError };
+ }
+
+ /** The note a continued attempt gets: it keeps the work, it does not redo it. */
+ function continuationNote(ending, words) {
+ return `\n\n---\n## Continuation (written by openvibe.codes)\n\nAn earlier attempt${ending.agent ? ` on ${nameOf(ending.agent)}` : ''} ${words}. Keep the work it finished, do not redo it, and continue from where it stopped. If the task is already complete, check it and report that.\n`;
+ }
+
+ /** The note a new agent gets: what changed, what it said, where it stopped. */
+ function handoverNote(ending, result, words) {
+ const d = result && typeof result === 'object' ? result : {};
+ const files = Array.isArray(d.files) ? d.files.map((f) => f && f.path).filter(Boolean).slice(0, 20) : [];
+ const summary = clip(d.summary || '', 1200);
+ return `\n\n---\n## Hand-over (written by openvibe.codes)\n\n${nameOf(ending.agent)} ${words}. Before it stopped it:\n`
+ + `- changed: ${files.length ? files.join(', ') : 'nothing yet'}\n`
+ + `- said: ${summary || 'nothing'}\n`
+ + `- stopped at: ${words}\n\n`
+ + 'Keep the work it did, do not redo it, and finish the task.\n';
+ }
+
+ /**
+ * What happens after an attempt ends: the engine decides, this turns the decision into a plan.
+ *
+ * @param {object} what
+ * @param {object} what.task the task row
+ * @param {object[]} what.attempts every attempt of the task, in order
+ * @param {object} what.ending the attempt that ended
+ * @param {object} [what.result] the job's result (what it changed and what it said)
+ * @returns {object} the plan: an outcome ('done', 'unfinished', 'stopped') or a follow-up attempt
+ */
+ function decide({ task, attempts, ending, result = null }) {
+ const ctx = routing();
+ const upto = (attempts || []).filter((a) => Number(a.n) <= Number(ending.n));
+ const facts = factsOf(result);
+ const engineAttempts = upto.map((a) => engineAttempt(a, upto, a.id === ending.id ? facts : null));
+ const out = chain.outcome(engineAttempts, ctx);
+ const step = out.step || { action: 'none' };
+ const reason = ending.reason || out.reason || 'error';
+ const words = wordsOf(reason);
+ const continues = upto.filter((a) => a.kind === 'continue').length;
+
+ if (step.action === 'done') return { outcome: 'done', note: 'the task finished' };
+ if (step.action === 'stop') return { outcome: 'stopped', note: 'the task was stopped on purpose' };
+ if (step.action === 'wait') return { outcome: null, note: null };
+
+ if (step.action === 'wrap-up') {
+ const hit = step.hit || 'steps';
+ return {
+ kind: 'wrap-up', agent: ending.agent, model: ending.model, session: ending.session || null,
+ prompt: String(task.prompt || '') + budget.wrapUpPrompt(hit),
+ why: `one last turn to land the work after the ${hit} limit`,
+ };
+ }
+
+ if (step.action === 'continue') {
+ if (continues >= maxContinues) {
+ return { outcome: 'unfinished', note: `the task was carried on ${continues} times already; it needs a person` };
+ }
+ return {
+ kind: 'continue', agent: ending.agent, model: ending.model, session: step.session || ending.session || null,
+ prompt: String(task.prompt || '') + continuationNote(ending, words),
+ why: `carried on in the same conversation: ${words}`,
+ };
+ }
+
+ if (step.action === 'restart') {
+ return {
+ kind: 'restart', agent: ending.agent, model: ending.model, session: null,
+ prompt: String(task.prompt || '') + continuationNote(ending, words),
+ why: `started again with a fresh session: ${words}`,
+ };
+ }
+
+ if (step.action === 'switch') {
+ const choice = nextAgent(task.prompt, ctx, runnerOf(ending.agent));
+ const agent = choice ? choice.agent : (step.runner ? HARNESS_OF[step.runner] : null);
+ if (!agent) return { outcome: 'unfinished', note: `no other agent could take it: ${step.why || words}` };
+ return {
+ kind: 'switch', agent, model: choice ? choice.model : (step.model || null), session: null,
+ prompt: String(task.prompt || '') + handoverNote(ending, result, words),
+ why: `${nameOf(agent)} took over: ${words}`,
+ };
+ }
+
+ return { outcome: 'unfinished', note: `${step.why || 'the task stopped'}. What is left is in the session.` };
+ }
+
+ /** An attempt as the API and the page show it. */
+ function present(a) {
+ if (!a) return null;
+ return {
+ id: a.id,
+ n: Number(a.n) || 0,
+ agent: a.agent || null,
+ model: a.model || null,
+ machine: a.machine || null,
+ job_id: a.job_id || null,
+ session: a.session || null,
+ state: a.state,
+ reason: a.reason || null,
+ why: a.why || null,
+ kind: a.kind || 'first',
+ cost_usd: a.cost_usd == null ? null : Number(a.cost_usd),
+ tokens: a.tokens == null ? null : Number(a.tokens),
+ steps: a.steps == null ? null : Number(a.steps),
+ started_at: iso(a.started_at),
+ finished_at: iso(a.finished_at),
+ created_at: iso(a.created_at),
+ end_why: a.reason ? wordsOf(a.reason) : null,
+ };
+ }
+
+ return {
+ STATES, REASONS, KINDS, REASON_TEXT, RUNNER_OF, HARNESS_OF,
+ list, get, byJob, create, bind, end, decide, present, routing, nextAgent, nameOf, wordsOf,
+ };
+}
+
+module.exports = { createAttempts, RUNNER_OF, HARNESS_OF, STATES, REASONS, KINDS, REASON_TEXT, ERROR_TEXT };
@@ -4,10 +4,13 @@
@@ -4,10 +4,13 @@
* Coding tasks on a person's own machine (run_tasks, plan T16): what the /run pages (server/http/run.js) and the
* Coding tasks on a person's own machine (run_tasks, plan T16): what the /run pages (server/http/run.js) and the
* service API (server/http/tasks-api.js, codes.task.*) share.
* service API (server/http/tasks-api.js, codes.task.*) share.
*
*
- * createRunTasks({ store, run, config, log, sessions, outbox })
+ * createRunTasks({ store, run, config, log, sessions, outbox, attempts })
* start({ subject, project, args, via, key }) → the row: queued (Run took the job) or failed (Run's reason)
* start({ subject, project, args, via, key }) → the row: queued (Run took the job) or failed (Run's reason)
* refresh(row, { force }) → the row, read from Run only as a fallback after 10 quiet minutes
* refresh(row, { force }) → the row, read from Run only as a fallback after 10 quiet minutes
* applyEvent(row, event) → a run.job.* delivery: state, timings, result, event, in one tx
* applyEvent(row, event) → a run.job.* delivery: state, timings, result, event, in one tx
+ * afterEnd(taskId) → the engine's decision after an attempt ended
+ * continueTask(row, { note, agent }) → the next attempt now (a switch when an agent is named)
+ * stopTask(row) → stop the running attempt; nothing carries it on
* latestProject(subject) → the project of the person's latest task, or null
* latestProject(subject) → the project of the person's latest task, or null
* present(row) → codes.task@1
* present(row) → codes.task@1
* TERMINAL, STATE_LABEL, STALE_MS
* TERMINAL, STATE_LABEL, STALE_MS
@@ -15,8 +18,15 @@
@@ -15,8 +18,15 @@
* Every state change writes its codes.task.* outbox row in the same transaction; a terminal state also back-fills
* Every state change writes its codes.task.* outbox row in the same transaction; a terminal state also back-fills
* the session's missing lines and marks it complete. A task's result comes from the person's machine: present() passes
* the session's missing lines and marks it complete. A task's result comes from the person's machine: present() passes
* on only the contract's fields, each checked and cut to its limit.
* on only the contract's fields, each checked and cut to its limit.
+ *
+ * Stage C2: every Run job is one attempt on the task's timeline (server/domain/attempts.js). When an attempt ends,
+ * the engine decides what happens next — done, a continue in the same session, a switch to another agent, a wrap-up or
+ * an unfinished task — and this module carries the decision out. One ending starts at most one follow-up: the
+ * follow-up's follow_up_of is unique, so two racing ticks never carry one job on twice.
*/
*/
const { RunError } = require('./run-client');
const { RunError } = require('./run-client');
+const { chain, budget } = require('../../engine');
+const { parseArgs } = require('../../harness/job');
const TERMINAL = new Set(['succeeded', 'failed', 'cancelled', 'expired']);
const TERMINAL = new Set(['succeeded', 'failed', 'cancelled', 'expired']);
const STATE_LABEL = {
const STATE_LABEL = {
@@ -85,12 +95,60 @@ function errorOf(job) {
@@ -85,12 +95,60 @@ function errorOf(job) {
return `${job.error.code || 'run.error'}${job.error.detail ? `: ${job.error.detail}` : ''}`.slice(0, 1000);
return `${job.error.code || 'run.error'}${job.error.detail ? `: ${job.error.detail}` : ''}`.slice(0, 1000);
}
}
-function createRunTasks({ store, run, config, log = console, sessions = null, outbox = null }) {
+/** The machine a job ran on: the worker's device id, or the hosted sandbox. */
+function machineOf(job) {
+ if (!job || typeof job !== 'object') return null;
+ const node = job.placement && typeof job.placement === 'object' ? job.placement.node : null;
+ if (typeof node === 'string' && node) return node.slice(0, 128);
+ return typeof job.machine === 'string' && job.machine ? job.machine.slice(0, 128) : null;
+}
+
+/** The attempt's own state for a Run job's state. */
+const ATTEMPT_STATE = { queued: 'queued', placed: 'queued', running: 'running', succeeded: 'done', failed: 'failed', cancelled: 'stopped', expired: 'failed' };
+
+/** The reason an attempt ended: the engine's own words (engine/chain.js reason), plus the two it reads from the text. */
+function reasonOf(state, text) {
+ if (state === 'cancelled') return 'stopped';
+ if (state === 'expired') return 'capacity';
+ const s = String(text || '');
+ if (/went quiet|hung|stale/i.test(s)) return 'hung';
+ if (/prompt is too long|context length|maximum context|context window|too many tokens|input too long/i.test(s)) return 'context';
+ const why = chain.reason(s);
+ return why === 'unknown' ? 'error' : why;
+}
+
+/** What an attempt spent: its cost, its tokens, its steps and the session a resume needs (from the job's result log). */
+function usageOf(result) {
+ const d = asObject(result) || {};
+ const log = Array.isArray(d.log) ? d.log : [];
+ let tokens = 0; let steps = 0; let session = null; let model = null;
+ for (const line of log) {
+ if (!line || typeof line !== 'object') continue;
+ if (line.t === 'usage') tokens += (Number(line.input_tokens) || 0) + (Number(line.output_tokens) || 0);
+ else if (line.t === 'tool') steps++;
+ else if (line.t === 'agent') {
+ if (!session && line.session) session = String(line.session);
+ if (!model && line.model) model = String(line.model);
+ }
+ }
+ return {
+ cost_usd: typeof d.cost_usd === 'number' && Number.isFinite(d.cost_usd) ? d.cost_usd : null,
+ tokens: tokens || null,
+ steps: steps || null,
+ session: session || null,
+ model: model || null,
+ };
+}
+
+function createRunTasks({ store, run, config, log = console, sessions = null, outbox = null, attempts = null }) {
const db = store.db;
const db = store.db;
const who = (row) => ({ subject: row.subject, project: row.project_id });
const who = (row) => ({ subject: row.subject, project: row.project_id });
// Lazily, so a caller that only reads tasks does not need the session/outbox modules (and there is no cycle).
// Lazily, so a caller that only reads tasks does not need the session/outbox modules (and there is no cycle).
const sessionLines = sessions || require('./session-lines').createSessionLines({ store, config, log });
const sessionLines = sessions || require('./session-lines').createSessionLines({ store, config, log });
const taskOutbox = outbox || require('../events/outbox').createOutbox({ store, log });
const taskOutbox = outbox || require('../events/outbox').createOutbox({ store, log });
+ const attemptStore = attempts || require('./attempts').createAttempts({ store, config, log });
+ /** An attempt's state when it ended: a limit reached is a budget-stop, never a plain failure. */
+ const ENDED = new Set(['done', 'failed', 'stopped', 'budget-stopped', 'interrupted']);
/**
/**
* One state change: update the row, write its codes.task.* outbox row in the same transaction, and on a terminal
* One state change: update the row, write its codes.task.* outbox row in the same transaction, and on a terminal
@@ -121,36 +179,79 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -121,36 +179,79 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
[current.id, state, result == null ? null : JSON.stringify(result), error, startedAt, finishedAt, store.now()],
[current.id, state, result == null ? null : JSON.stringify(result), error, startedAt, finishedAt, store.now()],
);
);
if (changed) await taskOutbox.emitTask(updated, current.state);
if (changed) await taskOutbox.emitTask(updated, current.state);
+ // Every Run job is one attempt: the attempt of this job follows the job.
+ const attempt = await attemptStore.byJob(current.job_id);
if (TERMINAL.has(updated.state) && !updated.session_complete_at) {
if (TERMINAL.has(updated.state) && !updated.session_complete_at) {
- await sessionLines.backfillAndComplete(updated, updated.result);
+ await sessionLines.backfillAndComplete(updated, updated.result, attempt ? Number(attempt.n) : null);
+ }
+ if (attempt) {
+ if (TERMINAL.has(state)) {
+ const usage = usageOf(updated.result);
+ const why = reasonOf(state, updated.error);
+ const hit = budgetHit(attempt, usage, why);
+ await attemptStore.end(attempt.id, {
+ state: hit ? 'budget-stopped' : (ATTEMPT_STATE[state] || 'failed'),
+ reason: hit ? 'budget' : why,
+ cost_usd: usage.cost_usd, tokens: usage.tokens, steps: usage.steps, session: usage.session,
+ model: usage.model, machine: machineOf(next.job),
+ finished_at: updated.finished_at || store.iso(),
+ });
+ } else {
+ await attemptStore.bind(attempt.id, {
+ state: ATTEMPT_STATE[state] || 'queued', started_at: updated.started_at || null, machine: machineOf(next.job),
+ });
+ }
}
}
return updated;
return updated;
});
});
}
}
+ /**
+ * The limit an attempt reached, if any (engine/budget.js): steps, minutes or dollars. A provider's own limit
+ * (usage limit, signed out) is not the owner's budget and never buys a wrap-up turn.
+ */
+ function budgetHit(attempt, usage, reason) {
+ if (!['error', 'hung', 'context'].includes(reason)) return null;
+ const limits = budget.limitsFor(config, attempt.agent, null);
+ if (!limits.steps && !limits.minutes && !limits.usd) return null;
+ const started = attempt.started_at ? Date.parse(String(attempt.started_at)) : 0;
+ const finished = attempt.finished_at ? Date.parse(String(attempt.finished_at)) : 0;
+ const minutes = finished > started ? (finished - started) / 60000 : null;
+ return budget.evaluate({ steps: usage.steps || 0, minutes, usd: usage.cost_usd }, limits).hit;
+ }
+
async function start({ subject, project, args, via = null, key = null, actorRef = null }) {
async function start({ subject, project, args, via = null, key = null, actorRef = null }) {
const id = store.newId('tsk');
const id = store.newId('tsk');
const at = store.now();
const at = store.now();
const actorRefValue = actorRef && typeof actorRef === 'object' && !Array.isArray(actorRef) ? JSON.stringify(actorRef) : null;
const actorRefValue = actorRef && typeof actorRef === 'object' && !Array.isArray(actorRef) ? JSON.stringify(actorRef) : null;
+ let inserted;
await store.tx(async () => {
await store.tx(async () => {
- const inserted = await db.one(`INSERT INTO run_tasks (id, subject, project_id, repo, ref, prompt, agent, push, pull_request, state, via, idempotency_key, actor_ref, created_at, updated_at)
+ inserted = await db.one(`INSERT INTO run_tasks (id, subject, project_id, repo, ref, prompt, agent, push, pull_request, state, via, idempotency_key, actor_ref, created_at, updated_at)
VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, 'queued', $10, $11, $12::jsonb, $13, $13) RETURNING *`,
VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, 'queued', $10, $11, $12::jsonb, $13, $13) RETURNING *`,
[id, subject, project, args.repo, args.ref, args.prompt, args.agent, !!args.push, !!(args.push && args.push.pull_request), via, key, actorRefValue, at]);
[id, subject, project, args.repo, args.ref, args.prompt, args.agent, !!args.push, !!(args.push && args.push.pull_request), via, key, actorRefValue, at]);
await taskOutbox.emitTask(inserted, null);
await taskOutbox.emitTask(inserted, null);
return inserted;
return inserted;
});
});
+ // The task's first attempt: every Run job is one attempt, and this one is the task as it was asked.
+ const first = await attemptStore.create({
+ task: inserted, agent: args.agent || null, model: args.model || null,
+ why: args.agent ? `the first try, with ${attemptStore.nameOf(args.agent)}` : 'the first try, with the agent that fits the task',
+ kind: 'first', state: 'queued',
+ });
let job;
let job;
try {
try {
job = await run.submit({ subject, project, args, key: id });
job = await run.submit({ subject, project, args, key: id });
} catch (e) {
} catch (e) {
const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;
const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;
+ await attemptStore.end(first.attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });
return await store.tx(async () => {
return await store.tx(async () => {
- const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, finished_at = $3::timestamptz, updated_at = $4
- WHERE id = $1 RETURNING *`, [id, why, store.iso(), store.now()]);
+ const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, outcome = 'unfinished', outcome_note = $3, finished_at = $4::timestamptz, updated_at = $5
+ WHERE id = $1 RETURNING *`, [id, why, `the task could not start: ${why}`, store.iso(), store.now()]);
await taskOutbox.emitTask(failed, 'queued');
await taskOutbox.emitTask(failed, 'queued');
return failed;
return failed;
});
});
}
}
+ await attemptStore.bind(first.attempt.id, { job_id: job.id, state: 'queued' });
return await store.tx(async () => {
return await store.tx(async () => {
const state = STATES.has(job.state) ? job.state : 'queued';
const state = STATES.has(job.state) ? job.state : 'queued';
const startedAt = job.timings && job.timings.started_at ? job.timings.started_at : null;
const startedAt = job.timings && job.timings.started_at ? job.timings.started_at : null;
@@ -165,7 +266,7 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -165,7 +266,7 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
async function applyJob(row, job) {
async function applyJob(row, job) {
if (!job || !STATES.has(job.state)) return row;
if (!job || !STATES.has(job.state)) return row;
const result = TERMINAL.has(job.state) ? (asObject(job.result) || null) : undefined;
const result = TERMINAL.has(job.state) ? (asObject(job.result) || null) : undefined;
- return await recordState(row, {
+ const updated = await recordState(row, {
state: job.state,
state: job.state,
result,
result,
error: TERMINAL.has(job.state) ? errorOf(job) : undefined,
error: TERMINAL.has(job.state) ? errorOf(job) : undefined,
@@ -173,6 +274,8 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -173,6 +274,8 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
finishedAt: job.timings && job.timings.finished_at,
finishedAt: job.timings && job.timings.finished_at,
job,
job,
});
});
+ // The attempt ended: the engine decides, and the row carries the outcome (or the next attempt).
+ return (await afterEnd(updated.id)) || updated;
}
}
/**
/**
@@ -180,7 +283,15 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -180,7 +283,15 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
* for a missing delivery, not page-view polling). force reads now, for a cancel or an explicit check.
* for a missing delivery, not page-view polling). force reads now, for a cancel or an explicit check.
*/
*/
async function refresh(row, { force = false } = {}) {
async function refresh(row, { force = false } = {}) {
- if (!row.job_id || TERMINAL.has(row.state)) return row;
+ if (!row.job_id) return row;
+ if (TERMINAL.has(row.state)) {
+ // The task has ended. Its decision (the outcome, or the attempt that follows) is made once, when the
+ // attempt ended; if that failed, it is made here, as the fallback reads Run after 10 quiet minutes.
+ if (!row.outcome && (force || store.now() - Number(row.updated_at) >= STALE_MS)) {
+ return (await afterEnd(row.id)) || row;
+ }
+ return row;
+ }
if (!force && store.now() - Number(row.updated_at) < STALE_MS) return row;
if (!force && store.now() - Number(row.updated_at) < STALE_MS) return row;
let job;
let job;
try { job = await run.read(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] task ${row.id}: ${e.message}`); return row; }
try { job = await run.read(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] task ${row.id}: ${e.message}`); return row; }
@@ -201,6 +312,160 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -201,6 +312,160 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
});
});
}
}
+ /**
+ * An attempt ended: the engine decides what happens next (server/domain/attempts.js) and this carries the
+ * decision out — a follow-up attempt, or the task's outcome in plain words.
+ *
+ * Calling it twice for one ending changes nothing: the follow-up's follow_up_of is unique, so the second call
+ * finds the attempt the first one wrote and starts no second job (the portal's 10-30 bug, where two racing ticks
+ * carried one job on twice).
+ */
+ async function afterEnd(taskId) {
+ const row = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [taskId]);
+ if (!row) return null;
+ if (!TERMINAL.has(row.state)) return row;
+ if (row.outcome === 'stopped') return row; // a person stopped it: nothing carries it on
+ const attempts = await attemptStore.list(row.id);
+ const ending = attempts[attempts.length - 1];
+ if (!ending || !ENDED.has(ending.state)) return row;
+ const plan = await attemptStore.decide({ task: row, attempts, ending, result: row.result });
+ if (plan.outcome) {
+ return await db.one('UPDATE run_tasks SET outcome = $2, outcome_note = $3, updated_at = $4 WHERE id = $1 RETURNING *',
+ [row.id, plan.outcome, plan.note, store.now()]);
+ }
+ if (!plan.kind) return row;
+ const done = await startFollowUp(row, plan, ending);
+ return done.row || row;
+ }
+
+ /**
+ * The next attempt: a Run job for the task with the engine's plan in it. The attempt is written first and the job
+ * second, so the fence (follow_up_of) is in place before anything is started.
+ */
+ async function startFollowUp(row, plan, ending) {
+ const { attempt, created } = await attemptStore.create({
+ task: row, agent: plan.agent || null, model: plan.model || null, session: plan.session || null,
+ why: plan.why || null, kind: plan.kind || 'manual', followUpOf: ending.id, state: 'queued',
+ });
+ // Someone else already started the follow-up of this ending: no second job.
+ if (!created) return { attempt, created: false, row };
+ let args;
+ try {
+ args = followUpArgs(row, plan);
+ } catch (e) {
+ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });
+ const failed = await db.one(`UPDATE run_tasks SET outcome = 'unfinished', outcome_note = $2, updated_at = $3 WHERE id = $1 RETURNING *`,
+ [row.id, `the next attempt could not be prepared: ${e.message}`, store.now()]);
+ return { attempt, created: true, row: failed, error: e.message };
+ }
+ let job;
+ try {
+ job = await run.submit({ subject: row.subject, project: row.project_id, args, key: attempt.id });
+ } catch (e) {
+ const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;
+ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });
+ const failed = await db.one(`UPDATE run_tasks SET outcome = 'unfinished', outcome_note = $2, updated_at = $3 WHERE id = $1 RETURNING *`,
+ [row.id, `the next attempt could not start: ${why}`, store.now()]);
+ return { attempt, created: true, row: failed, error: why };
+ }
+ await attemptStore.bind(attempt.id, { job_id: job.id, state: 'queued' });
+ const updated = await db.one(
+ `UPDATE run_tasks SET job_id = $2, state = 'queued', outcome = NULL, outcome_note = NULL, error = NULL,
+ finished_at = NULL, updated_at = $3 WHERE id = $1 RETURNING *`,
+ [row.id, job.id, store.now()]);
+ await taskOutbox.emitTask(updated, row.state);
+ return { attempt, created: true, job, row: updated };
+ }
+
+ /**
+ * The next attempt now: the owner carries the task on (with a note), or switches to another agent. The task's row
+ * is locked while the attempt is written, so two presses start one attempt and not two.
+ */
+ async function continueTask(row, { note = null, agent = null } = {}) {
+ if (!row || !row.id) return { error: 'no such task' };
+ const attempts = await attemptStore.list(row.id);
+ const ending = attempts[attempts.length - 1];
+ if (!ending) return { error: 'the task has no attempt to carry on' };
+ const text = note ? String(note).trim().slice(0, 4000) : '';
+ const plan = agent
+ ? {
+ kind: 'switch', agent, model: null, session: null,
+ prompt: `${row.prompt}\n\n---\n## Hand-over (written by openvibe.codes)\n\n${attemptStore.nameOf(agent)} takes over from ${attemptStore.nameOf(ending.agent)}, which ${attemptStore.wordsOf(ending.reason || 'error')}. Keep the work that was done, do not redo it, and finish the task.\n`,
+ why: `${attemptStore.nameOf(agent)} took over at the owner's request`,
+ }
+ : {
+ kind: 'manual', agent: ending.agent, model: ending.model, session: null,
+ prompt: `${row.prompt}${text ? `\n\n---\n## A note from the owner\n\n${text}\n` : ''}`,
+ why: text ? `carried on by the owner: ${text.slice(0, 200)}` : 'carried on by the owner',
+ };
+ let attempt = null;
+ const locked = await store.tx(async () => {
+ const current = await db.maybe('SELECT * FROM run_tasks WHERE id = $1 FOR UPDATE', [row.id]);
+ if (!current) return { error: 'no such task' };
+ if (!TERMINAL.has(current.state)) return { error: 'the task is still running' };
+ if (current.outcome === 'stopped') return { error: 'the task was stopped on purpose' };
+ const made = await attemptStore.create({
+ task: current, agent: plan.agent, model: plan.model, session: null, why: plan.why, kind: plan.kind, state: 'queued',
+ });
+ attempt = made.attempt;
+ await db.one(`UPDATE run_tasks SET state = 'queued', outcome = NULL, outcome_note = NULL, error = NULL, job_id = NULL, finished_at = NULL, updated_at = $2 WHERE id = $1 RETURNING *`,
+ [row.id, store.now()]);
+ return { row: current };
+ });
+ if (locked.error) return { error: locked.error };
+ let job;
+ try {
+ job = await run.submit({ subject: row.subject, project: row.project_id, args: followUpArgs(row, plan), key: attempt.id });
+ } catch (e) {
+ const why = e instanceof RunError ? `${e.code}: ${e.message}` : e.message;
+ await attemptStore.end(attempt.id, { state: 'failed', reason: 'error', finished_at: store.iso() });
+ const failed = await db.one(`UPDATE run_tasks SET state = 'failed', error = $2, outcome = 'unfinished', outcome_note = $3, finished_at = $4::timestamptz, updated_at = $5 WHERE id = $1 RETURNING *`,
+ [row.id, why, `the next attempt could not start: ${why}`, store.iso(), store.now()]);
+ return { error: why, row: failed, attempt };
+ }
+ await attemptStore.bind(attempt.id, { job_id: job.id, state: 'queued' });
+ const updated = await db.one('UPDATE run_tasks SET job_id = $2, updated_at = $3 WHERE id = $1 RETURNING *', [row.id, job.id, store.now()]);
+ await taskOutbox.emitTask(updated, row.state);
+ return { attempt, job, row: updated };
+ }
+
+ /** The job's arguments for a follow-up: the task's own repository, the plan's prompt and the session to resume. */
+ function followUpArgs(row, plan) {
+ return parseArgs(JSON.stringify({
+ repo: row.repo, ref: row.ref || null, prompt: plan.prompt || row.prompt,
+ agent: plan.agent || null, permission: 'edit',
+ ...(row.push ? { push: { pull_request: !!row.pull_request } } : {}),
+ ...(plan.session ? { resume: plan.session } : {}),
+ }));
+ }
+
+ /**
+ * Stop the running attempt: the outcome is stopped and nothing carries it on. The task's row is locked, so a stop
+ * and a follow-up never race.
+ */
+ async function stopTask(row) {
+ if (!row || !row.id) return { error: 'no such task' };
+ const attempts = await attemptStore.list(row.id);
+ const running = attempts.slice().reverse().find((a) => !ENDED.has(a.state));
+ if (row.job_id && !TERMINAL.has(row.state)) {
+ try { await run.cancel(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] stop ${row.id}: ${e.message}`); }
+ }
+ const stopped = await store.tx(async () => {
+ const current = await db.maybe('SELECT * FROM run_tasks WHERE id = $1 FOR UPDATE', [row.id]);
+ if (!current) return { error: 'no such task' };
+ if (TERMINAL.has(current.state) && current.outcome) return { error: 'the task has already ended', row: current };
+ const updated = await db.one(
+ `UPDATE run_tasks SET state = 'cancelled', outcome = 'stopped', outcome_note = $2,
+ finished_at = COALESCE(finished_at, $3::timestamptz), updated_at = $4 WHERE id = $1 RETURNING *`,
+ [row.id, 'the task was stopped on purpose', store.iso(), store.now()]);
+ await taskOutbox.emitTask(updated, current.state);
+ return { row: updated };
+ });
+ if (stopped.error) return stopped;
+ if (running) await attemptStore.end(running.id, { state: 'stopped', reason: 'stopped', finished_at: store.iso() });
+ return { row: stopped.row, attempt: running || null };
+ }
+
async function latestProject(subject) {
async function latestProject(subject) {
const row = await db.maybe('SELECT project_id FROM run_tasks WHERE subject = $1 ORDER BY created_at DESC LIMIT 1', [subject]);
const row = await db.maybe('SELECT project_id FROM run_tasks WHERE subject = $1 ORDER BY created_at DESC LIMIT 1', [subject]);
return row ? row.project_id : null;
return row ? row.project_id : null;
@@ -215,7 +480,10 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
@@ -215,7 +480,10 @@ function createRunTasks({ store, run, config, log = console, sessions = null, ou
});
});
}
}
- return { start, refresh, applyJob, applyEvent, recordState, latestProject, present, who };
+ return {
+ start, refresh, applyJob, applyEvent, recordState, afterEnd, continueTask, stopTask,
+ latestProject, present, who, attempts: attemptStore,
+ };
}
}
-module.exports = { createRunTasks, presentResult, TERMINAL, STATE_LABEL, STATES, RANK, STALE_MS, sessionLogOf, errorOf, asObject };
+module.exports = { createRunTasks, presentResult, TERMINAL, STATE_LABEL, STATES, RANK, STALE_MS, sessionLogOf, errorOf, asObject, ATTEMPT_STATE, reasonOf, usageOf };
@@ -98,7 +98,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -98,7 +98,7 @@ function createSessionLines({ store, config = {}, log = console }) {
const retentionDays = Math.max(1, Number(config.sessions && config.sessions.retentionDays) || 90);
const retentionDays = Math.max(1, Number(config.sessions && config.sessions.retentionDays) || 90);
/** Insert lines in order; an existing (task_id, seq) wins, so a replay never duplicates. → rows inserted. */
/** Insert lines in order; an existing (task_id, seq) wins, so a replay never duplicates. → rows inserted. */
- async function append(taskId, lines) {
+ async function append(taskId, lines, attemptN = null) {
if (!taskId || !Array.isArray(lines) || !lines.length) return 0;
if (!taskId || !Array.isArray(lines) || !lines.length) return 0;
const clean = [];
const clean = [];
for (const line of lines) {
for (const line of lines) {
@@ -114,12 +114,12 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -114,12 +114,12 @@ function createSessionLines({ store, config = {}, log = console }) {
const values = [];
const values = [];
const params = [];
const params = [];
chunk.forEach((line, j) => {
chunk.forEach((line, j) => {
- const b = j * 5;
- values.push(`($${b + 1}, $${b + 2}, $${b + 3}::timestamptz, $${b + 4}, $${b + 5}::jsonb)`);
- params.push(taskId, line.seq, line.at, line.kind, JSON.stringify(line.body));
+ const b = j * 6;
+ values.push(`($${b + 1}, $${b + 2}, $${b + 3}::timestamptz, $${b + 4}, $${b + 5}::jsonb, $${b + 6})`);
+ params.push(taskId, line.seq, line.at, line.kind, JSON.stringify(line.body), attemptN);
});
});
inserted += await db.exec(
inserted += await db.exec(
- `INSERT INTO codes_session_lines (task_id, seq, at, kind, body) VALUES ${values.join(', ')}
+ `INSERT INTO codes_session_lines (task_id, seq, at, kind, body, attempt_n) VALUES ${values.join(', ')}
ON CONFLICT (task_id, seq) DO NOTHING`, params);
ON CONFLICT (task_id, seq) DO NOTHING`, params);
}
}
return inserted;
return inserted;
@@ -127,7 +127,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -127,7 +127,7 @@ function createSessionLines({ store, config = {}, log = console }) {
/** All rows of a task in stream order. */
/** All rows of a task in stream order. */
async function allRows(taskId) {
async function allRows(taskId) {
- return await db.many('SELECT seq, at, kind, body FROM codes_session_lines WHERE task_id = $1 ORDER BY seq ASC', [taskId]);
+ return await db.many('SELECT seq, at, kind, body, attempt_n FROM codes_session_lines WHERE task_id = $1 ORDER BY seq ASC', [taskId]);
}
}
/**
/**
@@ -135,7 +135,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -135,7 +135,7 @@ function createSessionLines({ store, config = {}, log = console }) {
* into the job's JSON lines; state and end frames become their own lines. The stream's seq is the base, so the
* into the job's JSON lines; state and end frames become their own lines. The stream's seq is the base, so the
* same event replayed after a reconnect maps to the same rows.
* same event replayed after a reconnect maps to the same rows.
*/
*/
- function recorder(taskId) {
+ function recorder(taskId, { attemptN = null } = {}) {
let frameBuffer = '';
let frameBuffer = '';
let outputBuffer = '';
let outputBuffer = '';
let fallbackSeq = 0;
let fallbackSeq = 0;
@@ -196,7 +196,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -196,7 +196,7 @@ function createSessionLines({ store, config = {}, log = console }) {
frameBuffer = `${frameBuffer}${String(text == null ? '' : text)}`.replace(/\r\n/g, '\n');
frameBuffer = `${frameBuffer}${String(text == null ? '' : text)}`.replace(/\r\n/g, '\n');
const out = [];
const out = [];
drainFrames(out);
drainFrames(out);
- if (out.length) await append(taskId, out);
+ if (out.length) await append(taskId, out, attemptN);
},
},
async flush() {
async flush() {
const out = [];
const out = [];
@@ -208,7 +208,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -208,7 +208,7 @@ function createSessionLines({ store, config = {}, log = console }) {
out.push({ seq: (lastBase || 0) + PARTIAL_LINE_INDEX, at: store.iso(), kind: line.kind, body: line.body });
out.push({ seq: (lastBase || 0) + PARTIAL_LINE_INDEX, at: store.iso(), kind: line.kind, body: line.body });
outputBuffer = '';
outputBuffer = '';
}
}
- if (out.length) await append(taskId, out);
+ if (out.length) await append(taskId, out, attemptN);
},
},
ended: () => ended,
ended: () => ended,
};
};
@@ -219,7 +219,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -219,7 +219,7 @@ function createSessionLines({ store, config = {}, log = console }) {
* JSON, so a reconnect or a second result never adds a line twice; the new seqs sit after the streamed output and
* JSON, so a reconnect or a second result never adds a line twice; the new seqs sit after the streamed output and
* before the terminal state/end line.
* before the terminal state/end line.
*/
*/
- async function backfill(row, result) {
+ async function backfill(row, result, attemptN = null) {
if (!row || !row.id) return 0;
if (!row || !row.id) return 0;
const rows = await allRows(row.id);
const rows = await allRows(row.id);
const harness = rows.filter((r) => r && r.body && typeof r.body === 'object' && !Array.isArray(r.body)
const harness = rows.filter((r) => r && r.body && typeof r.body === 'object' && !Array.isArray(r.body)
@@ -242,7 +242,7 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -242,7 +242,7 @@ function createSessionLines({ store, config = {}, log = console }) {
const at = atIso(row.finished_at) || (row.updated_at ? iso(row.updated_at) : store.iso());
const at = atIso(row.finished_at) || (row.updated_at ? iso(row.updated_at) : store.iso());
return await append(row.id, missing.map((body, i) => ({
return await append(row.id, missing.map((body, i) => ({
seq: start + i, at, kind: typeof body.t === 'string' ? body.t : 'text', body,
seq: start + i, at, kind: typeof body.t === 'string' ? body.t : 'text', body,
- })));
+ })), attemptN);
}
}
async function complete(taskId) {
async function complete(taskId) {
@@ -252,8 +252,8 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -252,8 +252,8 @@ function createSessionLines({ store, config = {}, log = console }) {
);
);
}
}
- async function backfillAndComplete(row, result) {
- const inserted = await backfill(row, result);
+ async function backfillAndComplete(row, result, attemptN = null) {
+ const inserted = await backfill(row, result, attemptN);
await complete(row.id);
await complete(row.id);
return inserted;
return inserted;
}
}
@@ -276,11 +276,11 @@ function createSessionLines({ store, config = {}, log = console }) {
@@ -276,11 +276,11 @@ function createSessionLines({ store, config = {}, log = console }) {
let source;
let source;
if (rows.length) {
if (rows.length) {
source = 'session';
source = 'session';
- lines = rows.filter((r) => Number(r.seq) > from).map((r) => ({ seq: Number(r.seq), at: r.at, kind: r.kind, body: r.body }));
+ lines = rows.filter((r) => Number(r.seq) > from).map((r) => ({ seq: Number(r.seq), at: r.at, kind: r.kind, body: r.body, attempt_n: r.attempt_n == null ? null : Number(r.attempt_n) }));
} else {
} else {
source = 'result.log';
source = 'result.log';
const at = row.created_at ? iso(row.created_at) : store.iso();
const at = row.created_at ? iso(row.created_at) : store.iso();
- lines = (sessionLogOf(row) || []).map((body, i) => ({ seq: i + 1, at, kind: typeof body.t === 'string' ? body.t : 'text', body }))
+ lines = (sessionLogOf(row) || []).map((body, i) => ({ seq: i + 1, at, kind: typeof body.t === 'string' ? body.t : 'text', body, attempt_n: 1 }))
.filter((line) => line.seq > from);
.filter((line) => line.seq > from);
}
}
const page = lines.slice(0, max);
const page = lines.slice(0, max);
@@ -77,7 +77,7 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountS
@@ -77,7 +77,7 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountS
let job = null;
let job = null;
if (TERMINAL.has(state)) job = await run.read(jobId, tasks.who(row));
if (TERMINAL.has(state)) job = await run.read(jobId, tasks.who(row));
- return await store.tx(async () => {
+ const applied = await store.tx(async () => {
const inserted = await db.maybe(
const inserted = await db.maybe(
`INSERT INTO codes_events_inbox (event_id, event_type, task_id, outcome, processed_at)
`INSERT INTO codes_events_inbox (event_id, event_type, task_id, outcome, processed_at)
VALUES ($1, $2, $3, $4, $5::timestamptz) ON CONFLICT (event_id) DO NOTHING RETURNING event_id`,
VALUES ($1, $2, $3, $4, $5::timestamptz) ON CONFLICT (event_id) DO NOTHING RETURNING event_id`,
@@ -90,6 +90,13 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountS
@@ -90,6 +90,13 @@ function createEventsConsumer({ config, store, tasks, run, accountData, accountS
await db.exec('UPDATE codes_events_inbox SET outcome = $2 WHERE event_id = $1', [event.event_id, state]);
await db.exec('UPDATE codes_events_inbox SET outcome = $2 WHERE event_id = $1', [event.event_id, state]);
return { duplicate: false, outcome: state, task_id: row.id };
return { duplicate: false, outcome: state, task_id: row.id };
});
});
+ // The attempt ended: the engine decides what happens next. It runs outside the transaction, because a
+ // follow-up starts a Run job. A failure here is logged, not fatal: the delivery is applied, and the decision
+ // is made again by the fallback (the read of Run after 10 quiet minutes).
+ if (!applied.duplicate) {
+ try { await tasks.afterEnd(applied.task_id); } catch (err) { log.warn(`[Codes] task ${applied.task_id} decision failed: ${err.message}`); }
+ }
+ return applied;
}
}
const router = express.Router();
const router = express.Router();
@@ -5,7 +5,8 @@
@@ -5,7 +5,8 @@
* GET /run the form, the person's projects and their recent tasks (a sign-in prompt for a guest)
* GET /run the form, the person's projects and their recent tasks (a sign-in prompt for a guest)
* POST /run start one: a run_tasks row and an OpenVibe Run job for the person (server/domain/run-client.js)
* POST /run start one: a run_tasks row and an OpenVibe Run job for the person (server/domain/run-client.js)
* POST /run/project create a project for the runs when the person has none they may run jobs in
* POST /run/project create a project for the runs when the person has none they may run jobs in
- * GET /run/tasks/:id one task: live while it runs, then the agents' summary, the changed files and the patch
+ * GET /run/tasks/:id one task: its attempts (the timeline), live while it runs, then the agents' summary,
+ * the changed files and the patch
* GET /run/tasks/:id/events its live lines, proxied from Run's job stream (server-sent events)
* GET /run/tasks/:id/events its live lines, proxied from Run's job stream (server-sent events)
* POST /run/tasks/:id/cancel stop it
* POST /run/tasks/:id/cancel stop it
*
*
@@ -26,6 +27,9 @@ const DELEGABLE = new Set(['owner', 'admin']);
@@ -26,6 +27,9 @@ const DELEGABLE = new Set(['owner', 'admin']);
const MACHINES_URL = 'https://openvibe.bot/machines';
const MACHINES_URL = 'https://openvibe.bot/machines';
const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;
const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;
+/** The task's outcome in the words a person reads. */
+const OUTCOME_TEXT = { done: 'Done', unfinished: "Didn't finish", stopped: 'Stopped by you' };
+
/** A repository URL as a heading: the host small, then owner/name, which may break after a slash but never mid-word. */
/** A repository URL as a heading: the host small, then owner/name, which may break after a slash but never mid-word. */
function repoName(url) {
function repoName(url) {
const [host, ...path] = String(url).replace(/^https:\/\//, '').split('/');
const [host, ...path] = String(url).replace(/^https:\/\//, '').split('/');
@@ -182,6 +186,24 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat
@@ -182,6 +186,24 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat
</section>`;
</section>`;
}
}
+ /** The task's timeline: every attempt and why it happened. The data-* hooks are for the design of stage C4. */
+ async function timelineHtml(row) {
+ const attempts = await taskStore.attempts.list(row.id);
+ if (!attempts.length) return '';
+ const outcome = row.outcome
+ ? html`<p class="run-outcome" data-outcome="${row.outcome}"><b>${OUTCOME_TEXT[row.outcome] || row.outcome}</b>${row.outcome_note ? html` ${row.outcome_note}` : ''}</p>`
+ : '';
+ return html`<section class="run-timeline" data-timeline data-attempts="${attempts.length}">
+<h2>The attempts</h2>
+<ol class="run-attempts">${attempts.map((a) => html`<li class="run-attempt" data-attempt="${a.n}" data-attempt-id="${a.id}" data-state="${a.state}" data-reason="${a.reason || ''}" data-kind="${a.kind || 'first'}" data-agent="${a.agent || ''}">
+<p class="run-attempt-head"><b>Attempt ${a.n}</b>${a.agent ? html` · ${a.agent}` : ''} <span class="run-pill run-pill-${a.state}">${a.state}</span></p>
+<p class="run-attempt-why" data-why>${a.why || ''}</p>
+${a.reason ? html`<p class="run-attempt-end" data-end-why>${taskStore.attempts.wordsOf(a.reason)}</p>` : ''}
+</li>`)}</ol>
+${outcome}
+</section>`;
+ }
+
r.get('/run/tasks/:id', async (req, res) => {
r.get('/run/tasks/:id', async (req, res) => {
if (!signedIn(req)) return res.redirect(302, `/auth/login?next=${encodeURIComponent(req.originalUrl)}`);
if (!signedIn(req)) return res.redirect(302, `/auth/login?next=${encodeURIComponent(req.originalUrl)}`);
let row = await own(req, req.params.id);
let row = await own(req, req.params.id);
@@ -197,6 +219,7 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat
@@ -197,6 +219,7 @@ ${typeof d.duration_ms === 'number' ? html`<p class="muted">${Math.round(d.durat
<h1 class="run-repo">${repoName(row.repo)}${row.ref ? html` <small>${row.ref}</small>` : ''}</h1>
<h1 class="run-repo">${repoName(row.repo)}${row.ref ? html` <small>${row.ref}</small>` : ''}</h1>
<p><span class="run-pill run-pill-${row.state}" data-state>${STATE_LABEL[row.state] || row.state}</span>${row.agent ? html` <span class="run-pill">${row.agent}</span>` : ''}</p>
<p><span class="run-pill run-pill-${row.state}" data-state>${STATE_LABEL[row.state] || row.state}</span>${row.agent ? html` <span class="run-pill">${row.agent}</span>` : ''}</p>
<pre tabindex="0" class="run-prompt">${row.prompt}</pre>
<pre tabindex="0" class="run-prompt">${row.prompt}</pre>
+${await timelineHtml(row)}
${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events" data-title="${title}" aria-live="polite"><h2>Live</h2><ol class="run-log" data-log></ol></section>
${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events" data-title="${title}" aria-live="polite"><h2>Live</h2><ol class="run-log" data-log></ol></section>
<form method="post" action="/run/tasks/${row.id}/cancel"><button class="run-btn" type="submit">Cancel</button></form>` : html`${session}${resultHtml(row, { session: !!(lines && lines.length) })}`}`;
<form method="post" action="/run/tasks/${row.id}/cancel"><button class="run-btn" type="submit">Cancel</button></form>` : html`${session}${resultHtml(row, { session: !!(lines && lines.length) })}`}`;
return page(req, res, { title: 'Task', body, sharedScripts: ['agent-session.js'] });
return page(req, res, { title: 'Task', body, sharedScripts: ['agent-session.js'] });
@@ -242,7 +265,9 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events
@@ -242,7 +265,9 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events
if (!resp.ok || !resp.body) return res.status(resp.status === 404 ? 404 : 503).end();
if (!resp.ok || !resp.body) return res.status(resp.status === 404 ? 404 : 503).end();
res.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-store', 'X-Accel-Buffering': 'no', Connection: 'keep-alive' });
res.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-store', 'X-Accel-Buffering': 'no', Connection: 'keep-alive' });
// The bytes go to the browser unchanged, and the same frames become codes_session_lines as they stream.
// The bytes go to the browser unchanged, and the same frames become codes_session_lines as they stream.
- const recorder = sessionStore.recorder(row.id);
+ const attempts = await taskStore.attempts.list(row.id);
+ const attemptN = attempts.length ? Number(attempts[attempts.length - 1].n) : null;
+ const recorder = sessionStore.recorder(row.id, { attemptN });
const decoder = new TextDecoder();
const decoder = new TextDecoder();
try {
try {
for await (const chunk of resp.body) {
for await (const chunk of resp.body) {
@@ -286,13 +311,8 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events
@@ -286,13 +311,8 @@ ${live ? html`<section class="run-live" data-events="/run/tasks/${row.id}/events
if (!sameOrigin(req)) return res.status(403).type('text/plain').send('Cross-site form posts are refused.\n');
if (!sameOrigin(req)) return res.status(403).type('text/plain').send('Cross-site form posts are refused.\n');
const row = await own(req, req.params.id);
const row = await own(req, req.params.id);
if (!row) return res.status(404).type('text/plain').send('No such task of yours.\n');
if (!row) return res.status(404).type('text/plain').send('No such task of yours.\n');
- if (row.job_id && !TERMINAL.has(row.state)) {
- try { await run.cancel(row.job_id, who(row)); } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }
- try {
- const fresh = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [row.id]);
- if (fresh) await taskStore.refresh(fresh, { force: true });
- } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }
- }
+ // The running attempt stops: the outcome is "stopped" and nothing carries it on.
+ try { await taskStore.stopTask(row); } catch (e) { log.warn(`[Codes] cancel ${row.id}: ${e.message}`); }
return res.redirect(303, `/run/tasks/${row.id}`);
return res.redirect(303, `/run/tasks/${row.id}`);
});
});
new file mode 100644
new file mode 100644
@@ -0,0 +1,151 @@
@@ -0,0 +1,151 @@
+'use strict';
+
+/**
+ * A task's event stream (GET /api/v1/tasks/:id/events): the session's lines in the order they happened, the attempt
+ * each line belongs to, and the attempts themselves (started, ended with why). The stream is resumable: a client
+ * that reconnects sends the last line's seq as Last-Event-ID and gets only what came after it. A heartbeat every 15
+ * seconds keeps the connection alive through a proxy that would otherwise drop it.
+ *
+ * createTaskEvents({ store, config, log, attempts })
+ * stream(req, res, row) the stream of one task
+ *
+ * The lines come from codes_session_lines (the session the machine streamed, back-filled from its final result) and
+ * the attempts from codes_attempts. Nothing here writes anything: the stream only reads.
+ */
+
+const { TERMINAL } = require('../domain/run-tasks');
+
+const POLL_MS = 1000;
+const PAGE = 500;
+const ENDED = ['done', 'failed', 'stopped', 'budget-stopped', 'interrupted'];
+
+function createTaskEvents({ store, config = {}, log = console, attempts = null }) {
+ const db = store.db;
+ const heartbeatMs = Math.max(1000, Number(config.sse && config.sse.heartbeatMs) || 15_000);
+ const attemptStore = attempts || require('../domain/attempts').createAttempts({ store, config, log });
+
+ /** One SSE frame. */
+ const frame = (event, data, id = null) => (id == null ? '' : `id: ${id}\n`) + `event: ${event}\ndata: ${JSON.stringify(data)}\n\n`;
+
+ const started = (a) => ({
+ type: 'attempt_started', attempt_n: Number(a.n), id: a.id, agent: a.agent || null, model: a.model || null,
+ why: a.why || null, state: 'queued', at: a.created_at ? new Date(a.created_at).toISOString() : null,
+ });
+ const ended = (a) => ({
+ type: 'attempt_ended', attempt_n: Number(a.n), id: a.id, agent: a.agent || null, model: a.model || null,
+ why: a.why || null, state: a.state, reason: a.reason || null,
+ end_why: a.reason ? attemptStore.wordsOf(a.reason) : null,
+ cost_usd: a.cost_usd == null ? null : Number(a.cost_usd),
+ tokens: a.tokens == null ? null : Number(a.tokens),
+ steps: a.steps == null ? null : Number(a.steps),
+ at: a.finished_at ? new Date(a.finished_at).toISOString() : null,
+ });
+
+ /** The stream of one task: its lines, its attempts, a heartbeat, and an end when the task has ended. */
+ async function stream(req, res, row) {
+ res.writeHead(200, {
+ 'Content-Type': 'text/event-stream',
+ 'Cache-Control': 'no-store',
+ 'X-Accel-Buffering': 'no',
+ Connection: 'keep-alive',
+ });
+ if (res.flushHeaders) res.flushHeaders();
+
+ const asked = String(req.get('last-event-id') || (req.query && req.query.last_event_id) || '0').trim();
+ let lastSeq = /^\d+$/.test(asked) ? Number(asked) : 0;
+ let closed = false;
+ let busy = false;
+ let terminalAt = 0;
+
+ const write = (text) => {
+ if (closed || !text) return;
+ res.write(text);
+ };
+
+ // The attempts a client has not seen: from the attempt its last line belongs to.
+ const anchor = lastSeq
+ ? await db.maybe('SELECT attempt_n FROM codes_session_lines WHERE task_id = $1 AND seq = $2', [row.id, lastSeq])
+ : null;
+ const fromN = anchor && anchor.attempt_n ? Number(anchor.attempt_n) : 1;
+ const known = await attemptStore.list(row.id);
+ const sent = new Map(known.map((a) => [a.id, a.state]));
+ let out = '';
+ for (const a of known) {
+ if (Number(a.n) < fromN) continue;
+ out += frame('attempt', started(a));
+ if (ENDED.includes(a.state)) out += frame('attempt', ended(a));
+ }
+ write(out);
+
+ const sendLines = async () => {
+ const lines = await db.many(
+ 'SELECT seq, at, kind, body, attempt_n FROM codes_session_lines WHERE task_id = $1 AND seq > $2 ORDER BY seq ASC LIMIT $3',
+ [row.id, lastSeq, PAGE],
+ );
+ for (const line of lines) {
+ write(frame('line', {
+ seq: Number(line.seq),
+ at: line.at ? new Date(line.at).toISOString() : null,
+ kind: line.kind,
+ attempt_n: line.attempt_n == null ? null : Number(line.attempt_n),
+ body: line.body,
+ }, Number(line.seq)));
+ lastSeq = Math.max(lastSeq, Number(line.seq));
+ }
+ return lines.length;
+ };
+ await sendLines();
+
+ const tick = async () => {
+ if (closed || busy) return;
+ busy = true;
+ try {
+ const sentLines = await sendLines();
+ const fresh = await db.maybe('SELECT * FROM run_tasks WHERE id = $1', [row.id]);
+ const rows = await attemptStore.list(row.id);
+ let updates = '';
+ for (const a of rows) {
+ const was = sent.get(a.id);
+ if (was === a.state) continue;
+ sent.set(a.id, a.state);
+ if (was === undefined) updates += frame('attempt', started(a));
+ if (ENDED.includes(a.state)) updates += frame('attempt', ended(a));
+ }
+ write(updates);
+ if (fresh && TERMINAL.has(fresh.state) && !rows.some((a) => ['queued', 'running'].includes(a.state))) {
+ if (!terminalAt) terminalAt = Date.now();
+ // The task has ended and its outcome is known: the stream is done. A task whose decision is still
+ // being made (no outcome yet) keeps the stream a little longer, so a follow-up is not missed.
+ if (fresh.outcome || Date.now() - terminalAt > 10_000) {
+ closed = true;
+ clearInterval(poll);
+ clearInterval(beat);
+ res.end();
+ }
+ } else {
+ terminalAt = 0;
+ }
+ } catch (e) {
+ log.warn(`[Codes] task ${row.id} events: ${e.message}`);
+ } finally {
+ busy = false;
+ }
+ };
+
+ const poll = setInterval(tick, POLL_MS);
+ const beat = setInterval(() => write(': heartbeat\n\n'), heartbeatMs);
+ const stop = () => {
+ if (closed) return;
+ closed = true;
+ clearInterval(poll);
+ clearInterval(beat);
+ };
+ req.on('close', stop);
+ res.on('close', stop);
+ return undefined;
+ }
+
+ return { stream, heartbeatMs };
+}
+
+module.exports = { createTaskEvents };
@@ -7,19 +7,29 @@
@@ -7,19 +7,29 @@
* machine, in the project of their latest /run task.
* machine, in the project of their latest /run task.
*
*
* POST /tasks codes.task.start codes.task-start-request@1 → codes.task-start-result@1 (201, or 200 again)
* POST /tasks codes.task.start codes.task-start-request@1 → codes.task-start-result@1 (201, or 200 again)
- * GET /tasks/:id codes.task.read → codes.task@1
+ * GET /tasks/:id codes.task.read → codes.task@2 (the task as codes.task@1, with its attempts and
+ * its outcome added to it)
* POST /tasks/:id/cancel codes.task.cancel → codes.task@1
* POST /tasks/:id/cancel codes.task.cancel → codes.task@1
+ * GET /tasks/:id/events codes.task.read → the task's event stream (server-sent events)
+ * POST /tasks/:id/continue codes.task.start → the next attempt now (a switch when an agent is named)
+ * POST /tasks/:id/stop codes.task.cancel → the running attempt stops; nothing carries it on
*
*
* A Network service token for audience openvibe.codes (CODES_AUDIENCE), from a svc: principal holding the route's
* A Network service token for audience openvibe.codes (CODES_AUDIENCE), from a svc: principal holding the route's
- * capability. A service sees only the tasks it started, for that person; anything else is 404. The person's task
- * page (the codes.task@1 url) shows every task they own, whoever started it. Calls count against the person's own
- * per-actor budget (actor-limits.js), not the service's.
+ * capability. A service sees only the tasks it started, for that person; anything else is 404. The owner may read
+ * and act on their own tasks without a token. The person's task page (the codes.task@1 url) shows every task they
+ * own, whoever started it. Calls count against the person's own per-actor budget (actor-limits.js), not the service's.
+ *
+ * Stage C2: a task's timeline is its attempts. GET /tasks/:id answers codes.task@2: the task itself with the same
+ * fields it always had, and, added to them, its attempts and its outcome in plain words ("done", "unfinished" with
+ * what is left, or "stopped"). GET /tasks/:id/events replays the session's lines in order — resumable with
+ * Last-Event-ID, a heartbeat every 15 s — with the attempt each line belongs to.
*/
*/
const express = require('express');
const express = require('express');
const { serviceAuth, capabilities, http, validate } = require('openvibe-contracts');
const { serviceAuth, capabilities, http, validate } = require('openvibe-contracts');
const { asyncRouter } = require('./router');
const { asyncRouter } = require('./router');
const { parseArgs } = require('../../harness/job');
const { parseArgs } = require('../../harness/job');
const { createRunTasks, TERMINAL } = require('../domain/run-tasks');
const { createRunTasks, TERMINAL } = require('../domain/run-tasks');
+const { createTaskEvents } = require('./task-events');
const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;
const USER = /^usr_[0-9A-HJKMNP-TV-Z]{26}$/;
const TASK = /^tsk_[0-9A-HJKMNP-TV-Z]{26}$/;
const TASK = /^tsk_[0-9A-HJKMNP-TV-Z]{26}$/;
@@ -28,6 +38,7 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
@@ -28,6 +38,7 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
const r = asyncRouter();
const r = asyncRouter();
const db = store.db;
const db = store.db;
const taskStore = tasks || createRunTasks({ store, run, config, log });
const taskStore = tasks || createRunTasks({ store, run, config, log });
+ const taskEvents = createTaskEvents({ store, config, log, attempts: taskStore.attempts });
const problem = (req, res, status, code, detail) => http.sendProblem(res, status, code, { detail, ctx: req.ov });
const problem = (req, res, status, code, detail) => http.sendProblem(res, status, code, { detail, ctx: req.ov });
const json = express.json({ limit: '32kb' });
const json = express.json({ limit: '32kb' });
@@ -53,10 +64,32 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
@@ -53,10 +64,32 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
/** The task this service started for this person, or null. */
/** The task this service started for this person, or null. */
const mine = async (req) => {
const mine = async (req) => {
if (!TASK.test(String(req.params.id))) return null;
if (!TASK.test(String(req.params.id))) return null;
+ if (req.principal.owner) return db.maybe('SELECT * FROM run_tasks WHERE id = $1 AND subject = $2', [req.params.id, req.principal.subject]);
return db.maybe('SELECT * FROM run_tasks WHERE id = $1 AND subject = $2 AND via = $3', [req.params.id, req.principal.subject, req.principal.via]);
return db.maybe('SELECT * FROM run_tasks WHERE id = $1 AND subject = $2 AND via = $3', [req.params.id, req.principal.subject, req.principal.via]);
};
};
const byKey = (p, key) => db.maybe('SELECT * FROM run_tasks WHERE subject = $1 AND via = $2 AND idempotency_key = $3', [p.subject, p.via, key]);
const byKey = (p, key) => db.maybe('SELECT * FROM run_tasks WHERE subject = $1 AND via = $2 AND idempotency_key = $3', [p.subject, p.via, key]);
+ /** The owner (signed in), or a service token with the capability acting for the person in X-OV-Subject. */
+ const signedIn = (req) => req.viewer && req.viewer.kind === 'user' && req.viewer.subject;
+ const guardAny = (cap) => async (req, res, next) => {
+ if (String(req.headers.authorization || '').startsWith('Bearer ')) return guard(cap)(req, res, next);
+ if (signedIn(req)) {
+ req.principal = { sub: `user:${req.viewer.subject}`, subject: req.viewer.subject, via: null, owner: true };
+ return next();
+ }
+ return problem(req, res, 401, 'token.required', `${cap} needs a Network service token for ${config.serviceAudience}, or a signed-in owner`);
+ };
+
+ /** The task, its attempts and its outcome in plain words. The task's own fields stay at the top level. */
+ const timeline = async (row) => {
+ const attempts = await taskStore.attempts.list(row.id);
+ return {
+ ...taskStore.present(row),
+ attempts: attempts.map((a) => taskStore.attempts.present(a)),
+ outcome: { state: row.outcome || null, note: row.outcome_note || null },
+ };
+ };
+
r.post('/tasks', guard('codes.task.start'), actorLimits.budget('codes.task.start'), json, async (req, res) => {
r.post('/tasks', guard('codes.task.start'), actorLimits.budget('codes.task.start'), json, async (req, res) => {
const b = req.body;
const b = req.body;
// Actor back-reference: the contract does not carry these yet, so they are stripped before validation and stored
// Actor back-reference: the contract does not carry these yet, so they are stripped before validation and stored
@@ -102,10 +135,41 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
@@ -102,10 +135,41 @@ function createTasksApi({ config, store, keys, run, harnesses, actorLimits, task
return res.status(201).json({ task: taskStore.present(row), created: true });
return res.status(201).json({ task: taskStore.present(row), created: true });
});
});
- r.get('/tasks/:id', guard('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {
+ r.get('/tasks/:id', guardAny('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {
+ const row = await mine(req);
+ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');
+ return res.json(await timeline(await taskStore.refresh(row)));
+ });
+
+ /** The task's event stream: the session's lines in order, resumable with Last-Event-ID, plus the attempts. */
+ r.get('/tasks/:id/events', guardAny('codes.task.read'), actorLimits.reads('codes.read'), async (req, res) => {
+ const row = await mine(req);
+ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');
+ return await taskEvents.stream(req, res, row);
+ });
+
+ /** The next attempt now: the owner carries the task on, or hands it to another agent. */
+ r.post('/tasks/:id/continue', guardAny('codes.task.start'), actorLimits.budget('codes.task.start'), json, async (req, res) => {
+ const row = await mine(req);
+ if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');
+ const b = req.body || {};
+ const note = typeof b.note === 'string' ? b.note.slice(0, 4000) : null;
+ const agent = typeof b.agent === 'string' && b.agent ? b.agent : null;
+ if (agent && !harnesses.list().some((a) => a.id === agent)) return problem(req, res, 422, 'codes.agent_unknown', 'agent is one of GET /api/v1/harnesses');
+ const out = await taskStore.continueTask(row, { note, agent });
+ if (out.error) return problem(req, res, 409, 'codes.task_conflict', out.error);
+ log.info(`[Codes] task ${row.id} carried on (${agent ? `with ${agent}` : 'by the owner'})`);
+ return res.json(await timeline(out.row));
+ });
+
+ /** The running attempt stops: the outcome is "stopped" and nothing carries it on. */
+ r.post('/tasks/:id/stop', guardAny('codes.task.cancel'), actorLimits.budget('codes.task.start'), async (req, res) => {
const row = await mine(req);
const row = await mine(req);
if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');
if (!row) return problem(req, res, 404, 'codes.task_not_found', 'no such task of this service for this person');
- return res.json(taskStore.present(await taskStore.refresh(row)));
+ const out = await taskStore.stopTask(row);
+ if (out.error) return problem(req, res, 409, 'codes.task_conflict', out.error);
+ log.info(`[Codes] task ${row.id} stopped by ${req.principal.via || req.principal.subject}`);
+ return res.json(await timeline(out.row));
});
});
r.post('/tasks/:id/cancel', guard('codes.task.cancel'), actorLimits.budget('codes.task.start'), async (req, res) => {
r.post('/tasks/:id/cancel', guard('codes.task.cancel'), actorLimits.budget('codes.task.start'), async (req, res) => {
new file mode 100644
new file mode 100644
@@ -0,0 +1,288 @@
@@ -0,0 +1,288 @@
+'use strict';
+
+/**
+ * Stage C2: a task's attempts. Every Run job is one attempt; when an attempt ends the engine decides what happens
+ * next — done, a continue in the same conversation, a switch to another agent, a wrap-up after a limit, or an
+ * unfinished task — and one ending starts at most one follow-up (the portal's 10-30 bug, where two racing ticks
+ * carried one job on twice).
+ *
+ * The migration (an existing task becomes attempt 1), every engine decision, the fence under two concurrent endings,
+ * the stop fence, and the auth of every new route.
+ */
+const assert = require('assert');
+const { boot, check, done } = require('./helpers/boot');
+
+const TASK = /^tsk_[0-9A-HJKMNP-TV-Z]{26}$/;
+
+(async () => {
+ const calls = [];
+ const jobs = new Map();
+ let counter = 0;
+ const run = {
+ enabled: true,
+ async submit(o) {
+ counter++;
+ const id = `job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J${String(counter).padStart(2, '0')}`;
+ jobs.set(id, { id, state: 'queued', args: o.args });
+ calls.push(['submit', o]);
+ return { id, state: 'queued' };
+ },
+ async read(jobId) { return jobs.get(jobId) || { id: jobId, state: 'queued' }; },
+ async cancel(jobId, who) { calls.push(['cancel', jobId, who]); const j = jobs.get(jobId); if (j) j.state = 'cancelled'; return j; },
+ async streamUrl() { throw new Error('not used'); },
+ };
+ const t = await boot({ run, env: { CODES_TASK_STEPS: '2' } });
+ const db = t.ctx.store.db;
+ const ada = t.network.addUser('ada');
+ const bo = t.network.addUser('bo');
+ const CAPS = ['codes.task.start', 'codes.task.read', 'codes.task.cancel'];
+ const actor = t.network.mintService('svc:actor', 'openvibe.codes', CAPS);
+ const project = 'prj_01J8Z3Q4R5S6T7V8W9X0Y1Z2P1';
+
+ const call = async (method, path, { token = actor, subject = ada.subject, body, headers = {} } = {}) => {
+ const h = { Accept: 'application/json', ...headers, ...(token ? { Authorization: `Bearer ${token}` } : {}), ...(subject ? { 'X-OV-Subject': subject } : {}) };
+ if (body !== undefined) h['Content-Type'] = 'application/json';
+ const res = await fetch(`${t.base}/api/v1${path}`, { method, headers: h, body: body === undefined ? undefined : JSON.stringify(body) });
+ return { status: res.status, json: await res.json().catch(() => null) };
+ };
+
+ /** A task for Ada, started through the API (so it has its first attempt). */
+ const startTask = async (prompt = 'Fix the parser and add a test for it.') => {
+ const r = await call('POST', '/tasks', { body: { repo: 'https://github.com/octocat/Hello-World', prompt, idempotency_key: `act_${Math.random().toString(36).slice(2, 10)}` } });
+ assert.strictEqual(r.status, 201, JSON.stringify(r.json));
+ return r.json.task.id;
+ };
+
+ /** The attempt ends: the state change and then the engine's decision (what the events consumer does). */
+ const endAttempt = async (taskId, { state = 'failed', result = null, error = null } = {}) => {
+ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [taskId]);
+ const job = { id: row.job_id, state, result, error: error ? { code: 'run.job.worker_failed', detail: error } : null };
+ const updated = await t.ctx.tasks.applyEvent(row, { state, job, error });
+ await t.ctx.tasks.afterEnd(updated.id);
+ return await rowOf(updated.id);
+ };
+
+ const attemptsOf = async (taskId) => await db.many('SELECT * FROM codes_attempts WHERE task_id = $1 ORDER BY n ASC', [taskId]);
+ const rowOf = async (taskId) => await db.one('SELECT * FROM run_tasks WHERE id = $1', [taskId]);
+
+ try {
+ await check('the migration: every task that exists becomes attempt 1 of itself', async () => {
+ // A task from before the attempts (written straight into the table, as an old row would be).
+ const old = 'tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T9';
+ await db.exec(`INSERT INTO run_tasks (id, subject, project_id, job_id, repo, prompt, agent, state, created_at, updated_at)
+ VALUES ($1, $2, $3, 'job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J99', 'https://github.com/a/b', 'by hand', 'claude-code', 'succeeded', 1, 2)`, [old, ada.subject, project]);
+ const read = await t.ctx.attempts.list(old);
+ assert.strictEqual(read.length, 1, 'one attempt');
+ assert.strictEqual(Number(read[0].n), 1);
+ assert.strictEqual(read[0].state, 'done', 'the task succeeded, so its first attempt did');
+ assert.ok(read[0].why, 'the attempt says why it happened');
+ const shown = await t.get(`/api/v1/tasks/${old}`, { as: ada });
+ assert.strictEqual(shown.status, 200, shown.text.slice(0, 200));
+ assert.strictEqual(shown.json().attempts.length, 1);
+ assert.strictEqual(shown.json().attempts[0].n, 1);
+ });
+
+ await check('a task starts as one attempt, and the timeline says why', async () => {
+ const id = await startTask();
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 1);
+ assert.strictEqual(Number(attempts[0].n), 1);
+ assert.strictEqual(attempts[0].state, 'queued');
+ assert.strictEqual(attempts[0].kind, 'first');
+ assert.ok(attempts[0].job_id, 'the attempt carries the Run job');
+ const read = await call('GET', `/tasks/${id}`);
+ assert.strictEqual(read.json.id, id, 'the task at the top level, as codes.task@1 had it');
+ assert.strictEqual(read.json.state, 'queued');
+ assert.ok(read.json.url, 'the task carries its own fields');
+ assert.strictEqual(read.json.attempts.length, 1);
+ assert.strictEqual(read.json.outcome.state, null, 'a running task has no outcome yet');
+ });
+
+ await check('a finished attempt: the task is done, and nothing follows it', async () => {
+ const id = await startTask();
+ const updated = await endAttempt(id, { state: 'succeeded', result: { ok: true, summary: 'all done', files: [], log: [{ t: 'done', ok: true, text: 'all done' }] } });
+ assert.strictEqual(updated.outcome, 'done');
+ assert.strictEqual((await attemptsOf(id)).length, 1, 'no follow-up');
+ const read = await call('GET', `/tasks/${id}`);
+ assert.strictEqual(read.json.outcome.state, 'done');
+ assert.strictEqual(read.json.attempts[0].state, 'done');
+ });
+
+ await check('a crash: the same conversation continues, with the session to resume', async () => {
+ const id = await startTask();
+ const result = {
+ ok: false, summary: 'I will continue with the rest of it', files: [],
+ log: [{ t: 'agent', agent: 'claude-code', session: 'sess-42' }, { t: 'done', ok: false, text: 'I will continue with the rest of it' }],
+ };
+ await endAttempt(id, { state: 'failed', result, error: 'the machine went away' });
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 2, 'the task was carried on');
+ assert.strictEqual(attempts[1].kind, 'continue');
+ assert.strictEqual(attempts[1].session, 'sess-42', 'the session to resume');
+ assert.ok(/^carried on in the same conversation: /.test(attempts[1].why), attempts[1].why);
+ // The new attempt is a Run job of its own, with RESUME set to the session.
+ const submits = calls.filter((c) => c[0] === 'submit');
+ const last = submits[submits.length - 1][1];
+ assert.strictEqual(last.args.resume, 'sess-42');
+ assert.ok(last.args.prompt.includes('Continuation'), 'the prompt tells the agent to carry on');
+ assert.notStrictEqual(last.key, id, 'the follow-up has its own idempotency key');
+ });
+
+ await check('a limit: another agent takes over, with a hand-over note of what the last one did', async () => {
+ const id = await startTask();
+ const result = { ok: false, summary: 'I fixed the parser', files: [{ path: 'src/parse.js', added: 3, removed: 1 }], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] };
+ await endAttempt(id, { state: 'failed', result, error: 'usage limit reached' });
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 2);
+ assert.strictEqual(attempts[1].kind, 'switch');
+ assert.ok(attempts[1].agent && attempts[1].agent !== attempts[0].agent, `${attempts[1].agent} took over from ${attempts[0].agent}`);
+ assert.ok(/took over: /.test(attempts[1].why), attempts[1].why);
+ const submits = calls.filter((c) => c[0] === 'submit');
+ const last = submits[submits.length - 1][1];
+ assert.ok(last.args.prompt.includes('Hand-over'), 'the new agent gets a hand-over note');
+ assert.ok(last.args.prompt.includes('src/parse.js'), 'the note says what changed');
+ assert.ok(last.args.prompt.includes('I fixed the parser'), 'the note says what it said');
+ });
+
+ await check('a limit of the owner: one last turn to land the work', async () => {
+ const id = await startTask();
+ const result = {
+ ok: false, summary: 'half done', files: [],
+ log: [{ t: 'tool', id: 't1', name: 'Edit' }, { t: 'tool', id: 't2', name: 'Edit' }, { t: 'done', ok: false, text: 'maximum turns' }],
+ };
+ await endAttempt(id, { state: 'failed', result, error: 'the agent stopped' });
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 2);
+ assert.strictEqual(attempts[0].state, 'budget-stopped', 'a limit is not a failure');
+ assert.strictEqual(attempts[0].reason, 'budget');
+ assert.strictEqual(attempts[1].kind, 'wrap-up');
+ assert.ok(/one last turn to land the work/.test(attempts[1].why), attempts[1].why);
+ const submits = calls.filter((c) => c[0] === 'submit');
+ assert.ok(submits[submits.length - 1][1].args.prompt.includes('one short last turn'), 'the wrap-up prompt');
+ });
+
+ await check('the same failure twice: the task is unfinished, with what is left', async () => {
+ const id = await startTask();
+ const fail = { ok: false, summary: 'it broke', files: [], log: [{ t: 'done', ok: false, text: 'the tool call failed' }] };
+ await endAttempt(id, { state: 'failed', result: fail, error: 'the tool call failed' });
+ await endAttempt(id, { state: 'failed', result: fail, error: 'the tool call failed' });
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 2, 'a second opinion, and no more');
+ const row = await rowOf(id);
+ assert.strictEqual(row.outcome, 'unfinished');
+ assert.ok(row.outcome_note && row.outcome_note.length > 10, 'the outcome says what is left');
+ });
+
+ await check('one ending starts at most one follow-up, even when two endings race', async () => {
+ const id = await startTask();
+ const before = calls.filter((c) => c[0] === 'submit').length;
+ const row = await rowOf(id);
+ const job = { id: row.job_id, state: 'failed', result: { ok: false, summary: 'x', files: [], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] }, error: { code: 'run.job.worker_failed', detail: 'usage limit reached' } };
+ // Two ticks see the same ending at the same time (the portal's 10-30 bug).
+ await Promise.all([
+ t.ctx.tasks.applyEvent(row, { state: 'failed', job, error: 'usage limit reached' }),
+ t.ctx.tasks.applyEvent(row, { state: 'failed', job, error: 'usage limit reached' }),
+ ]);
+ await Promise.all([t.ctx.tasks.afterEnd(id), t.ctx.tasks.afterEnd(id)]);
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.length, 2, 'one follow-up, not two');
+ const submits = calls.filter((c) => c[0] === 'submit').length - before;
+ assert.strictEqual(submits, 1, 'one job started, not two');
+ });
+
+ await check('a stop fences the task: the outcome is stopped and nothing carries it on', async () => {
+ const id = await startTask();
+ const before = calls.filter((c) => c[0] === 'submit').length;
+ const stopped = await call('POST', `/tasks/${id}/stop`);
+ assert.strictEqual(stopped.status, 200, JSON.stringify(stopped.json));
+ assert.strictEqual(stopped.json.outcome.state, 'stopped');
+ assert.ok(calls.some((c) => c[0] === 'cancel'), 'the job was cancelled');
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts[0].state, 'stopped');
+ assert.strictEqual(attempts[0].reason, 'stopped');
+ // The machine's late report changes nothing.
+ await endAttempt(id, { state: 'failed', error: 'the machine went away' });
+ assert.strictEqual((await attemptsOf(id)).length, 1, 'nothing carried it on');
+ assert.strictEqual(calls.filter((c) => c[0] === 'submit').length - before, 0, 'no job was started');
+ assert.strictEqual((await rowOf(id)).outcome, 'stopped');
+ });
+
+ await check('the owner carries the task on, or hands it to another agent', async () => {
+ const id = await startTask();
+ await endAttempt(id, { state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] } });
+ const continued = await call('POST', `/tasks/${id}/continue`, { body: { note: 'Only the tests are left.' } });
+ assert.strictEqual(continued.status, 200, JSON.stringify(continued.json));
+ assert.strictEqual(continued.json.attempts.length, 2);
+ assert.strictEqual(continued.json.attempts[1].kind, 'manual');
+ assert.ok(/carried on by the owner/.test(continued.json.attempts[1].why), continued.json.attempts[1].why);
+ // While the task runs again, nobody carries it on a second time.
+ const again = await call('POST', `/tasks/${id}/continue`, { body: { agent: 'codex' } });
+ assert.strictEqual(again.status, 409, 'the task is running again');
+ });
+
+ await check('the auth of every new route: no token, no capability, another person', async () => {
+ const id = await startTask();
+ const noCap = t.network.mintService('svc:nocap', 'openvibe.codes', ['codes.other']);
+ const other = t.network.mintService('svc:other', 'openvibe.codes', CAPS);
+ // No token at all.
+ for (const [method, path] of [['GET', `/tasks/${id}`], ['GET', `/tasks/${id}/events`], ['POST', `/tasks/${id}/continue`], ['POST', `/tasks/${id}/stop`]]) {
+ const res = await call(method, path, { token: null });
+ assert.strictEqual(res.status, 401, `${method} ${path} without a token`);
+ }
+ // A service without the capability.
+ assert.strictEqual((await call('GET', `/tasks/${id}`, { token: noCap })).status, 403);
+ assert.strictEqual((await call('POST', `/tasks/${id}/continue`, { token: noCap, body: {} })).status, 403);
+ // Another person, and another service (which sees only the tasks it started).
+ assert.strictEqual((await call('GET', `/tasks/${id}`, { subject: bo.subject })).status, 404);
+ assert.strictEqual((await call('GET', `/tasks/${id}`, { token: other })).status, 404);
+ assert.strictEqual((await call('POST', `/tasks/${id}/stop`, { subject: bo.subject })).status, 404);
+ // A task that does not exist.
+ assert.strictEqual((await call('GET', '/tasks/tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T1')).status, 404);
+ assert.strictEqual((await call('GET', '/tasks/nope')).status, 404);
+ });
+
+ await check('the task page shows the timeline of attempts, each with its why', async () => {
+ const id = await startTask();
+ const result = { ok: false, summary: 'x', files: [], log: [{ t: 'done', ok: false, text: 'usage limit reached' }] };
+ await endAttempt(id, { state: 'failed', result, error: 'usage limit reached' });
+ const page = await t.get(`/run/tasks/${id}`, { as: ada });
+ assert.strictEqual(page.status, 200, page.text.slice(0, 300));
+ assert.ok(page.text.includes('data-timeline'), 'the timeline is there');
+ assert.ok(page.text.includes('data-attempt="1"') && page.text.includes('data-attempt="2"'), 'every attempt is a row');
+ assert.ok(page.text.includes('data-state="failed"'), 'the attempt says how it ended');
+ assert.ok(page.text.includes('data-why'), 'the attempt says why it happened');
+ assert.ok(/took over: /.test(page.text), 'the second attempt says who took over');
+ assert.ok(page.text.indexOf('data-timeline') < page.text.indexOf('data-session') || !page.text.includes('data-session'),
+ 'the timeline is above the session');
+ });
+
+ await check('a task is carried on twice, and then it needs a person', async () => {
+ const id = await startTask();
+ const crash = {
+ ok: false, summary: 'I will continue with the rest of it', files: [],
+ log: [{ t: 'agent', agent: 'claude-code', session: 'sess-1' }, { t: 'done', ok: false, text: 'I will continue with the rest of it' }],
+ };
+ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });
+ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });
+ const attempts = await attemptsOf(id);
+ assert.strictEqual(attempts.filter((a) => a.kind === 'continue').length, 2, 'two automatic continues');
+ await endAttempt(id, { state: 'failed', result: crash, error: 'the machine went away' });
+ const row = await rowOf(id);
+ assert.strictEqual(row.outcome, 'unfinished', 'the third one needs a person');
+ assert.ok(/carried on 2 times/.test(row.outcome_note), row.outcome_note);
+ assert.strictEqual((await attemptsOf(id)).length, 3, 'no third continue');
+ });
+
+ await check('the owner needs no token: their own task reads, continues and stops', async () => {
+ const id = await startTask();
+ const mine = await t.get(`/api/v1/tasks/${id}`, { as: ada });
+ assert.strictEqual(mine.status, 200, mine.text.slice(0, 200));
+ assert.strictEqual(mine.json().attempts.length, 1);
+ const theirs = await t.get(`/api/v1/tasks/${id}`, { as: bo });
+ assert.strictEqual(theirs.status, 404);
+ });
+ } finally {
+ await t.close();
+ }
+ done();
+})().catch((e) => { console.error(e); process.exit(1); });
new file mode 100644
new file mode 100644
@@ -0,0 +1,138 @@
@@ -0,0 +1,138 @@
+'use strict';
+
+/**
+ * A task's event stream (GET /api/v1/tasks/:id/events): the session's lines in the order they happened, the attempt
+ * each line belongs to, and the attempts themselves (started, ended with why). A client that reconnects sends the last
+ * line's seq as Last-Event-ID and gets only what came after it; a heartbeat keeps the connection alive.
+ */
+const assert = require('assert');
+const http = require('http');
+const { boot, check, done } = require('./helpers/boot');
+
+(async () => {
+ const calls = [];
+ let counter = 0;
+ const run = {
+ enabled: true,
+ async submit(o) {
+ counter++;
+ const id = `job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J${String(counter).padStart(2, '0')}`;
+ calls.push(['submit', o]);
+ return { id, state: 'queued' };
+ },
+ async read(jobId) { return { id: jobId, state: 'queued' }; },
+ async cancel(jobId) { return { id: jobId, state: 'cancelled' }; },
+ async streamUrl() { throw new Error('not used'); },
+ };
+ const t = await boot({ run, env: { CODES_SSE_HEARTBEAT_MS: '300' } });
+ const db = t.ctx.store.db;
+ const ada = t.network.addUser('ada');
+ const bo = t.network.addUser('bo');
+ const CAPS = ['codes.task.start', 'codes.task.read', 'codes.task.cancel'];
+ const actor = t.network.mintService('svc:actor', 'openvibe.codes', CAPS);
+
+ const call = async (method, path, { token = actor, subject = ada.subject, body, headers = {} } = {}) => {
+ const h = { Accept: 'text/event-stream', ...headers, ...(token ? { Authorization: `Bearer ${token}` } : {}), ...(subject ? { 'X-OV-Subject': subject } : {}) };
+ if (body !== undefined) h['Content-Type'] = 'application/json';
+ const res = await fetch(`${t.base}/api/v1${path}`, { method, headers: h, body: body === undefined ? undefined : JSON.stringify(body) });
+ return { status: res.status, text: await res.text() };
+ };
+
+ /** The stream, read until it ends (a finished task) or until the marker arrives. */
+ const stream = (path, { headers = {}, until = null, ms = 8000 } = {}) => new Promise((resolve, reject) => {
+ const req = http.get(`${t.base}${path}`, { headers }, (res) => {
+ let text = '';
+ res.setEncoding('utf8');
+ const timer = setTimeout(() => { req.destroy(); resolve({ status: res.statusCode, text }); }, ms);
+ res.on('data', (chunk) => {
+ text += chunk;
+ if (until && text.includes(until)) { clearTimeout(timer); req.destroy(); resolve({ status: res.statusCode, text }); }
+ });
+ res.on('end', () => { clearTimeout(timer); resolve({ status: res.statusCode, text }); });
+ });
+ req.on('error', (e) => (e.code === 'ECONNRESET' ? resolve({ status: 0, text: '' }) : reject(e)));
+ });
+
+ /** A task of Ada's with two attempts and three lines of session. */
+ const makeTask = async () => {
+ const r = await fetch(`${t.base}/api/v1/tasks`, {
+ method: 'POST',
+ headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${actor}`, 'X-OV-Subject': ada.subject },
+ body: JSON.stringify({ repo: 'https://github.com/octocat/Hello-World', prompt: 'Add a test for the date parser.', idempotency_key: `act_${Math.random().toString(36).slice(2, 10)}` }),
+ });
+ const json = await r.json();
+ const id = json.task.id;
+ await db.exec(`INSERT INTO codes_session_lines (task_id, seq, at, kind, body, attempt_n) VALUES
+ ($1, 1, now(), 'text', '{"t":"text","text":"first"}'::jsonb, 1),
+ ($1, 2, now(), 'text', '{"t":"text","text":"second"}'::jsonb, 1),
+ ($1, 3, now(), 'text', '{"t":"text","text":"third"}'::jsonb, 2)`, [id]);
+ return id;
+ };
+
+ try {
+ // Ada ran a task on /run before (so Codes knows her project).
+ await db.exec(`INSERT INTO run_tasks (id, subject, project_id, job_id, repo, prompt, state, created_at, updated_at)
+ VALUES ('tsk_01J8Z3Q4R5S6T7V8W9X0Y1Z2T1', $1, 'prj_01J8Z3Q4R5S6T7V8W9X0Y1Z2P1', 'job_01J8Z3Q4R5S6T7V8W9X0Y1Z2J00', 'https://github.com/a/b', 'by hand', 'succeeded', 1, 2)`, [ada.subject]);
+
+ await check('the stream: the lines in order, the attempt each belongs to, and the attempts themselves', async () => {
+ const id = await makeTask();
+ // The task ends, so the stream ends with it.
+ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [id]);
+ const job = { id: row.job_id, state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] }, error: null };
+ await t.ctx.tasks.applyEvent(row, { state: 'succeeded', job });
+ await t.ctx.tasks.afterEnd(id);
+ const s = await call('GET', `/tasks/${id}/events`);
+ assert.strictEqual(s.status, 200);
+ assert.ok(s.text.includes('event: line'), s.text.slice(0, 300));
+ const seqs = [...s.text.matchAll(/^id: (\d+)$/gm)].map((m) => Number(m[1]));
+ assert.deepStrictEqual(seqs.slice(0, 3), [1, 2, 3], 'the lines in order, each with its seq');
+ const bodies = [...s.text.matchAll(/^data: (.*)$/gm)].map((m) => JSON.parse(m[1]));
+ const lines = bodies.filter((b) => b.seq);
+ assert.deepStrictEqual(lines.slice(0, 3).map((l) => l.attempt_n), [1, 1, 2], 'each line says which attempt it belongs to');
+ assert.deepStrictEqual(lines.slice(0, 3).map((l) => l.body.text), ['first', 'second', 'third']);
+ const attempts = bodies.filter((b) => b.type);
+ assert.strictEqual(attempts[0].type, 'attempt_started');
+ assert.strictEqual(attempts[attempts.length - 1].type, 'attempt_ended');
+ assert.strictEqual(attempts[attempts.length - 1].state, 'done');
+ });
+
+ await check('a client that reconnects gets only what came after the line it last saw', async () => {
+ const id = await makeTask();
+ const row = await db.one('SELECT * FROM run_tasks WHERE id = $1', [id]);
+ const job = { id: row.job_id, state: 'succeeded', result: { ok: true, summary: 'done', files: [], log: [{ t: 'done', ok: true, text: 'done' }] }, error: null };
+ await t.ctx.tasks.applyEvent(row, { state: 'succeeded', job });
+ await t.ctx.tasks.afterEnd(id);
+ const s = await call('GET', `/tasks/${id}/events`, { headers: { 'Last-Event-ID': '2' } });
+ assert.strictEqual(s.status, 200);
+ const seqs = [...s.text.matchAll(/^id: (\d+)$/gm)].map((m) => Number(m[1]));
+ assert.deepStrictEqual(seqs[0], 3, 'only the line after the last one it saw');
+ assert.ok(!seqs.includes(1) && !seqs.includes(2), 'and not the ones before');
+ assert.ok(!s.text.includes('"first"') && !s.text.includes('"second"'), 'and not the ones before');
+ });
+
+ await check('the heartbeat keeps the connection alive', async () => {
+ const id = await makeTask();
+ const s = await stream(`/api/v1/tasks/${id}/events`, {
+ headers: { Authorization: `Bearer ${actor}`, 'X-OV-Subject': ada.subject },
+ until: ': heartbeat',
+ ms: 5000,
+ });
+ assert.ok(s.text.includes(': heartbeat'), s.text.slice(0, 200));
+ });
+
+ await check('the auth of the stream: no token, another person, another service', async () => {
+ const id = await makeTask();
+ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { token: null })).status, 401);
+ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { subject: bo.subject })).status, 404);
+ const other = t.network.mintService('svc:other', 'openvibe.codes', CAPS);
+ assert.strictEqual((await call('GET', `/tasks/${id}/events`, { token: other })).status, 404);
+ // The owner needs no token (the task runs, so the stream is read until the heartbeat).
+ const mine = await stream(`/api/v1/tasks/${id}/events`, { headers: { cookie: t.signIn(ada) }, until: ': heartbeat', ms: 5000 });
+ assert.strictEqual(mine.status, 200);
+ assert.ok(mine.text.includes('event: line'), mine.text.slice(0, 200));
+ });
+ } finally {
+ await t.close();
+ }
+ done();
+})().catch((e) => { console.error(e); process.exit(1); });
@@ -92,12 +92,24 @@ const { boot, check, done } = require('./helpers/boot');
@@ -92,12 +92,24 @@ const { boot, check, done } = require('./helpers/boot');
await db.exec('UPDATE run_tasks SET updated_at = updated_at - $2 WHERE id = $1', [task.id, 11 * 60_000]);
await db.exec('UPDATE run_tasks SET updated_at = updated_at - $2 WHERE id = $1', [task.id, 11 * 60_000]);
const r = await call('GET', `/tasks/${task.id}`);
const r = await call('GET', `/tasks/${task.id}`);
assert.strictEqual(r.status, 200, JSON.stringify(r.json));
assert.strictEqual(r.status, 200, JSON.stringify(r.json));
- const v = validate('codes.task@1', r.json);
+ // The read is the task (its own fields at the top level, as codes.task@1 had them), with its
+ // attempts and its outcome added to it: codes.task@2 (stage C2).
+ const v = validate('codes.task@2', r.json);
assert.ok(v.valid, JSON.stringify(v.errors));
assert.ok(v.valid, JSON.stringify(v.errors));
+ const v1doc = { ...r.json };
+ delete v1doc.attempts; delete v1doc.outcome;
+ const v1 = validate('codes.task@1', v1doc);
+ assert.ok(v1.valid, 'the task itself still answers codes.task@1');
assert.strictEqual(r.json.state, 'succeeded');
assert.strictEqual(r.json.state, 'succeeded');
+ assert.strictEqual(r.json.id, task.id, 'the task at the top level, not wrapped');
assert.strictEqual(r.json.result.summary.length, 4000);
assert.strictEqual(r.json.result.summary.length, 4000);
assert.deepStrictEqual(r.json.result.files, [{ path: 'README', added: 1, removed: 0 }]);
assert.deepStrictEqual(r.json.result.files, [{ path: 'README', added: 1, removed: 0 }]);
assert.deepStrictEqual(r.json.result.push, { pushed: false, reason: 'no token' });
assert.deepStrictEqual(r.json.result.push, { pushed: false, reason: 'no token' });
+ assert.strictEqual(r.json.attempts.length, 1, 'one attempt: the task as it was asked');
+ assert.strictEqual(r.json.attempts[0].n, 1);
+ assert.strictEqual(r.json.attempts[0].state, 'done');
+ assert.ok(r.json.attempts[0].why, 'the attempt says why it happened');
+ assert.strictEqual(r.json.outcome.state, 'done');
});
});
await check('another person, another service, or a task the person started by hand: 404', async () => {
await check('another person, another service, or a task the person started by hand: 404', async () => {