openvibers/openvibe.run

OpenVibe.Run

The job service of the network (openvibe.run). Design: OpenVibe.Contracts docs/adr/ADR-034-platform-north-star.md;
plan track T14.

Live site: https://openvibe.run · API docs: https://openvibe.services · For AI: https://openvibe.run/llms.txt
Run it: npm ci && npm run dev (Node 22 or newer) · Tests: npm test

Purpose

OpenVibe.Run accepts jobs, places each one on a node that advertises the runtime class it needs, and follows it to
its end. A job belongs to a project, is billed per wall-clock second it runs, and is visible only to the token
that project presents. Run owns the API, the project scope and the placement; OpenVibe.Bot owns the node link and the
per-second metering of the jobs it carries (plan T14 L1). The one exception is the external managed-browser provider
(plan T14 R4a): Run meters its jobs itself and posts their readings with a Billing token of its own, used for nothing
else.

This repository is at plan T14 steps 5 (R1b), 6 (R1c) and 7 (R3): the API is complete, and a queued job is
placed from the Fabric's offers and actually runs — the poller in server/jobs/poller.js (started by
server/index.js) sends it to OpenVibe.Bot, mirrors Bot's state and the node's stdout back, and settles it. Without
OV_OAUTH_CLIENT_SECRET Run cannot mint the Network token Bot requires, so the bridge is off and every job stays
queued.

How a job runs

1. Submit — POST /api/v1/jobs (capability run.job.submit). Run validates the request against
   run.job-create-request@1, mints the job_<ULID> id, stores the row queued and writes run.job.queued in the
   same transaction. Nothing leaves the process here: a submit never waits on Network or Bot.
2. Place — the poller (RUN_POLL_MS, default 1000 ms) reads the Fabric's offers for the job's class
   (GET {OV_NETWORK_URL}/api/v1/offers?kind=node), and openvibe-sdk/placement's plan() chooses the node the
   job's own requirements allow. Run records the node, provider and region on its row (placed_at included).
   No eligible offer: the job waits queued (a node may be briefly offline) and every tick tries again; if its
   ttl_ms runs out first it ends expired with run.job.unplaceable (ADR-036 section 2).
3. Send — Run hands the platform.job@1 body to Bot (POST {RUN_BOT_URL}/api/v1/jobs, capability
   bot.job.dispatch, audience openvibe.bot) and Bot gives it to the node over its device link. Bot's dispatch is
   idempotent by job id, so a retry after a crash or a timeout is safe.
4. Mirror — every tick mirrors each placed or running job from GET {RUN_BOT_URL}/api/v1/jobs/{id}: state,
   timings, exit, result and the metered wall clock (Bot is the meter; Run never posts a usage record for a job Bot carries). The state
   change and its run.job.started|succeeded|failed|cancelled|expired event are one transaction, and the job's
   stdout is carried into its SSE stream as output events. A job Bot answers 404 for after Run sent it fails
   run.job.worker_failed after RUN_DISPATCH_UNKNOWN_POLLS polls — it is never mirrored forever.
5. End — succeeded, failed, cancelled or expired, from Bot's job_exit (or from the cancel request and
   the ttl sweep). An end state never changes. Bot's answer is never trusted over Run's own row: it can neither
   rename the node nor the project of a job Run placed.

Owns

• The public site at openvibe.run (server/web/home.js, server/app.js): the front page in the OpenVibe Frame (what a
  job is and how it runs, the API in brief), the legal pages (openvibe-shared/legal), robots.txt, the sitemap,
  /llms.txt and /llms-full.txt, IndexNow (only with INDEXNOW_KEY: the key file and one ping at boot), the icon and the web manifest. It says only what works today; test/site.test.js checks it.
• The job API (server/api/v1.js, one capability per route — the table below) and the job's lifecycle record
  (run_jobs, migrations/0001_jobs.sql): state, placement, timings, exit, result and the metered wall clock.
• The project scope: a job's project_id and requester come from the caller's Network token
  (claims.project_id, claims.sub) and never from a request body; another project's job is 404, never 403.
• Project delegation (ADR-003 amendment 2026-10-12): a first-party service acting for a person (X-OV-Subject) may
  name a project with X-OV-Project. Run asks Network whether the person is an admin or owner of the unarchived
  project (network.project.member.read, a yes cached 60 s), then the person is the requester and the service is
  stored as via (migrations/0002_job_via.sql). The service submits, reads, cancels and gets a stream ticket for only
  the jobs it started for that person (another person's or another service's job is 404); list and a delegated Bearer
  stream are 403 project.not_delegated, and so is any other caller naming a project. Network unreachable is
  503 project.check_unavailable. Watch's run-job action and OpenVibe.Codes' coding jobs use it.
• The job_<ULID> id (minted here, passed unchanged to Bot as platform.job@1 id, so every metering reading of
  the job is run:<id>:<n>).
• Idempotent submit: one row per (project_id, idempotency_key), a repeat answering the first job, a different body
  with the same key refused 409 run.idempotency.conflict.
• The run.job.* events (server/jobs/events.js, run_events_outbox), written in the same transaction as the
  state change they describe.
• The job stream (server/jobs/stream.js): one ring buffer per open job, the SSE id being the job's own monotonic
  seq, plus the two-minute single-use stream ticket (server/jobs/ticket.js).
• The platform.workload-requirements@1 default ({ kind: run.<class>, mobility: job, latency_class: background, objective: balanced }, the contract's own default) and the project caps on a job's limits.
• Placement (plan T14 R3, server/dispatch/placement.js): the job's requirements are stated by Run and the choice
  among offers is openvibe-sdk/placement plan() — Run writes no placer of its own (ADR-036). The job's
  worker:<class> capability is what a candidate must advertise. A user-owned offer is a person's own machine
  (contracts 0.161.0): it is eligible only for that person's own job (requester user:usr_… equal to the offer's
  owner) whose requirements name user-owned. Network keeps such offers off the public list, so Run reads them,
  owners included, from GET /internal/resources?kind=node&trust=user-owned with its own client
  (network.registry.read); a user-owned offer on the public list is never used.
• The poller (server/jobs/poller.js): the ttl sweep, the send, the mirror, the cancel-on-Bot and the job's stdout
  into the stream.

Does not own

• The node link, the device credentials and the per-second metering: OpenVibe.Bot (bot.job.dispatch), which gives
  the job to a node advertising worker:<class> and meters every wall-clock second into platform.usage-sample@1
  readings keyed run:<job id>:<n>.
• Identity, tokens and projects (OpenVibe.Network); Run only verifies the tokens it is the audience of.
• The artifacts a job runs and the media its inputs come from (OpenVibe.Media; Run never fetches a caller-chosen
  URL).
• Money for the jobs Bot carries: Bot writes their usage records. Run writes only the external provider's
  (billing.usage.record, clamped to Run's own observed duration + 5 s).

Depends on

• PostgreSQL 18: run_jobs and the run_events_outbox; migrations/ applies on boot.
• Valkey (VALKEY_URL, optional): shared state between processes; unset, nothing is shared.
• OpenVibe.Network: the RS256 signing key (JWKS) every token is verified against; the offers Run places from
  (GET /api/v1/offers, public); Run's client run mints the events relay's service token, its Bot token and the
  token for project memberships (GET /internal/projects/:project/members/:subject, project delegation).
• OpenVibe.Events: run.job.* events go to events.event.publish; unset, they wait in run_events_outbox.
• OpenVibe.Bench (RUN_BENCH_URL, default https://bench.openvibe.services): a race's judgement is recorded as a
  bench.match@1 on POST /api/v1/matches (bench.judgement.cast, audience openvibe.bench). Bench has no
  matchmaker endpoint yet; matchmaker: true falls back to RUN_RACE_ENTRANTS.
• OpenVibe.Bot (RUN_BOT_URL, RUN_POLL_MS): the node link and the meter. Run sends each job to it
  (bot.job.dispatch, audience openvibe.bot, minted from OV_OAUTH_CLIENT_ID/OV_OAUTH_CLIENT_SECRET against
  OV_NETWORK_INTERNAL_URL and cached until 60 s before expiry) and mirrors its answer. Unset the client secret and
  the bridge is off: /api/ready skips the bot check and every job stays queued.
• openvibe-contracts v0.106.0, openvibe-sdk v0.74.0 and openvibe-shared v4.179.0 (package.json).

Capabilities and routes

Every route checks exactly the capability its manifest names (contracts/manifests/capabilities/run.job.*.json),
from the caller's service token. A Network user token names no project and holds no capability here, so it is
refused like any other token without the grant.

| Method and path | Capability | Request → answer |
|---------------------------------|------------------|--------------------------------------------------------|
| POST /api/v1/jobs | run.job.submit | run.job-create-request@1 → run.job-create-result@1 (201 created; 200, created:false on an idempotent repeat) |
| GET /api/v1/jobs/:id | run.job.read | common.no-body@1 → run.job-read-result@1 |
| GET /api/v1/jobs?state=&limit=&cursor= | run.job.list | run.job-list-query@1 → run.job-list-result@1 |
| POST /api/v1/jobs/:id/cancel | run.job.cancel | common.no-body@1 → run.job-read-result@1 |
| POST /api/v1/jobs/:id/stream/ticket | run.job.stream | common.no-body@1 → run.job-stream-ticket-result@1 |
| GET /api/v1/jobs/:id/stream?ticket= | run.job.stream | a ticket in the query, or a Bearer service token with the grant → text/event-stream of run.job-stream-event@1 |
| GET /api/v1/admin/jobs | run.job.admin | run.job-list-query@1 (with project_id) → run.job-list-result@1 |

run.job.admin additionally reads and cancels any project's job and is the only way to pass project_id to a list
(the manifest says so); it is internal, so only a first-party svc: token may hold it — an app, mod or agent token
with the grant is refused. Every request body and query string is validated against its released contract, and so is
every answer before it is sent; errors are RFC 9457 application/problem+json.

Checks

The checks scheduler (server/checks/, [docs/checks.md](docs/checks.md)) turns a matching GitHub workflow job into a
Run code job on a node and reports a GitHub check run. The owner creates the App once at /admin/checks (the
manifest flow); GitHub calls POST /webhooks/github; GET /api/v1/checks/agreement?repo= gives the owner the node vs
Actions agreement. Its secrets (the App key, webhook secret, client secret and installation tokens) are sealed with
RUN_SECRETS_KEY; RUN_CHECKS_PROJECT is the project a check job belongs to.

Sessions

A session (contracts 0.223.0, the Desks plan's D6) is a browser desk Run places for one person and OpenVibe.Bot leases
to it (bot.desk.session, server/sessions/). Unlike jobs, a session is a person's own: their Network token, or a
first-party svc: token holding the capability and naming them in X-OV-Subject (recorded as via).

| Method and path | Capability | Request → answer |
|-----------------------------------------|-----------------------|------------------|
| POST /api/v1/sessions | run.session.create | run.session-create-request@1 → run.session@1 (201; 200 on an idempotent repeat) |
| GET /api/v1/sessions?state=&limit=&cursor= | run.session.read | → run.session-list-result@1 (running ones by default) |
| GET /api/v1/sessions/:id | run.session.read | → run.session@1 |
| POST /api/v1/sessions/:id/commands | run.session.control | run.session-control-request@1 (screen, input, browser) → run.session-control-result@1 |
| POST /api/v1/sessions/:id/end | run.session.control | → run.session@1, ended |

• Placement. The request's own machine, or nowhere. Otherwise the hosts Bot lists for the person: online with
  room, in the trust classes the requirements allow (first-party, then community, by default; user-owned only when
  the requirements name it), most free desks first. A full host is skipped; with none, the session waits placing
  until its ttl_s, then ends unplaceable.
• Its life. ttl_s (default 3600) bounds it. idle_s (default 900) without a command ends it. Every
  RUN_SESSION_LEASE_CHECK_MS (30 s) Run asks Bot whether the desk is still leased: a host that dropped it ends the
  session host_lost. A person holds RUN_SESSIONS_PER_PERSON (3) at once.
• Metering. One platform.usage-sample@1 per minute of desk time (run:<session>:<minute>, operation
  browser.desk, the last one the remainder) in run_usage_outbox, posted to Billing by the usage relay.
• Events. run.session.started and run.session.ended, in the transaction that records each.

Jobs

• Submit. Idempotency-Key (or body.idempotency_key) makes a retry safe; with neither, Run mints its own key
  and returns it in the read result (auto:<ulid>). The row is unique per (project_id, idempotency_key) and that
  index is permanent, so a key is spent for good: a repeat always answers the first job, and the same key with
  different work is 409 — the contract promising that "within 24 h" is answered the same way, forever, rather than
  reused. limits are capped by the project's (RUN_LIMITS_*, 422 run.limits.exceeded) — the lifetime too, whether
  named or defaulted — args over 512 KiB are refused 413 run.args_too_large, and ttl_ms defaults to
  limits.wall_ms + 600000 exactly as the contract says.
• States. queued → placed → running → succeeded | failed | cancelled | expired. Run stores queued and, once
  the job is on a node, records placed itself; running and the end states arrive from Bot. An end state never
  changes, so a replay of a cancel is a no-op rather than a second event.
• Events. run.job.queued|started|succeeded|failed|cancelled|expired, payloads being the
  run.job-read-result@1 projection their contracts describe. The outbox row is keyed run:<job id>:<state> (one
  row per job and state, so a replay is one row); the published envelope carries a contract-valid evt_<ULID>
  document id, because events.event-envelope@1 pins that pattern and OpenVibe.Events validates it on publish.
• Stream. POST /jobs/:id/stream/ticket mints a two-minute single-use RS256 ticket for that one job (audience
  openvibe.run, typ run-stream, its own key — never Network's signing key, so it is never a session anywhere).
  GET /jobs/:id/stream?ticket=… replays what is kept (the last 1 MiB) and follows; a service that can send a header
  opens the same stream with a Bearer token holding run.job.stream for the job's project (the manifest allows
  both). Last-Event-ID resumes after the last seq seen; an ended job replays what is kept and closes with
  type: 'end'. Reconnecting never re-runs a job. The output events are the node's job_stdout chunks as Bot
  holds them (best effort: a chunk lost with the node's link is never resent, and a gap Bot dropped is dropped here
  rather than streamed twice). The ring is this process's memory (reclaimed once it ended or went idle, and rebuilt from the job's row when
  a stream opens again), so seq is this process's counter: a resume across a restart gets the current state rather
  than a gap, and output is best effort either way.
• The job record. run_jobs.job holds the platform.job@1 body Run will send (id, class, artifact, args,
  ttl_ms, limits, net, inputs) plus the fields platform.job@1 has no place for and 0001_jobs.sql has no column
  for: requirements (required by run.job-read-result@1 and by placement), placed_at (timings.placed_at) and
  cancel_requested (a cancel of a placed or running job). store.frameOf(record) builds the frame body from the
  platform.job@1 keys alone, and the dispatcher bridge validates it before anything is sent.
• Cancelling. A queued job was never sent anywhere: cancelled at once (exit stays null, usage stays 0). A
  placed or running job may be on a node this instant, so it is marked cancel_requested, Bot is asked for its
  job_cancel (once per job — Bot repeats it on every reconnect until the job ends — and the poller retries the
  request until it lands), and the job reaches cancelled when Bot says so. Run never answers cancelled for work a
  node may still be doing. Cancelling an ended job answers it as it is.
• Usage. usage.wall_ms and usage.seconds are Bot's metered wall clock (Bot writes the
  run:<job id>:<n> readings; Run never posts those). A running job therefore shows 0
  until its job_exit carries the wall clock: Run does not estimate a number it cannot back. A job the managed-browser
  provider ran is metered by Run: the provider's claimed wall_ms, never above Run's own observed duration (its send
  to the answer that showed the end) + 5 s — a larger claim is clamped and logged — queued as run:<job id>:<n>
  readings in run_usage_outbox with the end state and relayed to Billing (RUN_BILLING_URL; unset, they wait).
• Placement. The node is chosen per tick from the offers Network publishes
  (GET {OV_NETWORK_URL}/api/v1/offers?kind=node) by openvibe-sdk/placement plan(), on the job's own
  requirements; node_id, provider and region are recorded on Run's row. A class no node offers keeps the job
  queued until an offer appears or its ttl_ms ends it expired with run.job.unplaceable. Offers are cached for RUN_OFFERS_TTL_MS; a Network that does not answer
  is retried, never a reason to fail a job.

Races (the Arena on the one engine)

A race is the Arena's round (compare-arena-ecosystem.md §3): the same prompt, two to four entrants, one child job per
entrant on the same machine class (linux), the same share of the parent budget each, and no entrant named until a
person judges. The play is data — agents.play@1, id race, one role per side, one parallel phase, blind: true,
stop.when: person — listed by GET /api/v1/plays. Contracts 0.422.0 has no run.race@1, so the request and answer
are validated in server/races/service.js; the bench.match@1 the judgement records is validated before it is sent.

| Method and path | Capability | Request → answer |
|-----------------------------------|------------------|------------------|
| GET /api/v1/plays | run.job.list | → { plays: agents.play@1[] } |
| POST /api/v1/races | run.job.submit | { prompt, kind: chat\|code\|design, entrants: [2–4 bench.entrant ids] \| matchmaker: true, budget: { usd?, wall_ms?, cpu_ms?, mem_bytes? } } → 201 race |
| GET /api/v1/races/:id | run.job.read | each side's output so far (the job stream's ring), its state and, once ended, its receipts; names only after a judgement |
| POST /api/v1/races/:id/judgement | run.job.submit | { winner: a\|b\|tie\|both_bad } → 201 race with the names revealed |

• Who may start and judge. A person, through a first-party service (X-OV-Subject + X-OV-Project): Run records
  the person as the requester and the service as via. Only that person, through that service and project, may read
  or judge it; anyone else gets the same 404 run.race_not_found a missing race does.
• The split. budget is the round's total and each side gets floor(total / sides) of wall, CPU and memory
  (memory floored to bytes); usd is split and recorded, not metered by Run. A total too small for every side is
  422 run.limits.exceeded; the per-side limits then pass the project caps like any job.
• The child. class: linux, args: { play, role, race, kind, prompt, entrant, budget }, with
  job.parent = { type: 'race', id } (dropped by frameOf, so Bot never receives it). Placement, metering, events
  and the stream are the ordinary job path.
• The judgement. One bench.match@1, source arena, the shown order (a, b, …, randomised when the race is
  created), blind: true, each ended side's receipts, play: { id: 'race', run_id }, and one human_pairwise
  judgement whose by.id_hash is an HMAC of the person (RUN_RACE_HASH_KEY, falling back to the Network client
  secret). Bench records it before Run writes the judgement and reveals the names; a Bench that cannot answer leaves
  the race blind and retryable with the same match id.
• The matchmaker. Bench has no matchmaker endpoint yet (checked 2026-11-13). matchmaker: true picks two entrants
  of different model families at random from RUN_RACE_ENTRANTS; unset, it answers 422 run.race_matchmaker_unavailable and names the setting.

Tests

npm test              # every file in test/, PGlite (in-process PostgreSQL)
npm run test:pg       # the same on the production-shaped containers (openvibe-sdk scripts/test-services.sh up)
node test/jobs.test.js      # the job API against the released contracts
node test/dispatch.test.js  # the dispatcher bridge: offers → placement → Bot → mirror → stream, cancel, ttl
node test/app.test.js       # the skeleton: health, readiness, release.json, /metrics, the bridge's off state

test/helpers/app.js boots Run against an in-process OpenVibe.Network stub (its JWKS, its tokens, the token
endpoint Run mints from and the offers registry), so nothing needs the network, OpenVibe.Bot or a running service.
test/dispatch.test.js runs a fake OpenVibe.Bot as a real HTTP server that verifies the Network service token Run
mints for it (audience openvibe.bot, capability bot.job.dispatch), so the bridge is exercised over the wire
rather than through a fetch stub.

Security

• Every route checks one capability; the project and the requester come from the token, never the body.
• Another project's job is 404 run.job_not_found, never 403: an id's existence is never confirmed.
• The stream route accepts a valid, unexpired, unused ticket for that job, or a Bearer service token holding
  run.job.stream for the job's project — a token in the query string is refused like any other non-ticket.
• Bot is a trusted peer but never an authority: its answer cannot rename the node or the project of a job Run
  placed, cannot move a job's state backwards, and cannot make Run carry out work the caller did not ask for. Run's
  own row, its offers and the job's requirements decide.
• Run's Bot token is minted from OV_OAUTH_CLIENT_ID/SECRET, cached until 60 s before expiry, never logged and never
  put in an error; the job body is validated against platform.job@1 before it is sent.
• Limits are bounded by the contract and the project's caps; nothing is silently lowered.
• No secret is logged (no Authorization header, no ticket, no job output), and nothing starts at module load.
• See SECURITY.md to report a problem.

Deploy

On openvibe-ovh since 2026-10-07: unit openvibe-run on loopback :4920, release layout under /opt/openvibe.run,
env /etc/openvibe/run.env, database ov_run; deploy with sudo ovhost deploy run. First-party callers reach the job
API on loopback, and jobs are placed on people's own Bot nodes. openvibe.run is parked (a 302 to openvibe.network)
until Run has a product surface; the vhost below is installed then.

• deploy/nginx/openvibe.run.conf — the vhost (TLS, www → apex, rate limits, client IP from $remote_addr only,
  /internal/ and /metrics never proxied).
• deploy/systemd/openvibe-run.service — unit openvibe-run, WorkingDirectory=/opt/openvibe.run,
  EnvironmentFile=/etc/openvibe/run.env, PORT=4920.
• /api/health and /api/ready are for the loopback caller; /metrics refuses anything that arrived through a
  proxy.

Deploy files

| File | What it is |
|------|------------|
| deploy/nginx/openvibe.run.conf | The nginx vhost for openvibe.run (reference; deployed to /etc/nginx/sites-available/) |
| deploy/systemd/openvibe-run.service | The systemd unit (release layout /opt/openvibe.run) |
| .github/workflows/ci.yml | CI: the shared test job (syntax, tests, openvibe-contracts-check --service run) and the shared security job |
| migrations/0001_jobs.sql | run_jobs and run_events_outbox |
| migrations/0004_telemetry.sql | telemetry_samples: request telemetry (openvibe-sdk 0.46, platform.telemetry-sample@1) — each HTTP request aggregated by route template, method and status class, never the raw URL, with /shared files, static assets, health, readiness and /metrics not counted; once a minute one http.request sample per key, plus the active-requests, p95 latency and event-loop-lag gauges; rows older than 14 days deleted daily; wired in server/app.js (telemetryMiddleware, a no-op until init) and server/index.js (telemetry.init() at production start, retention, flush on stop) |
| STATUS.json | What this repository is, what it does today and where it runs |

<!-- versions:start -->

• openvibe-contracts: v0.422.0
• openvibe-sdk: v0.161.0
• openvibe-shared: v4.237.0

<!-- versions:end -->

Compute placement, pool autoscaling, rental caps and the operator view are documented in
[docs/compute-broker.md](docs/compute-broker.md). Rentals default to quote-only; keys and caps
are never inferred from available credentials.