Skip to content

Default Jobs

Instar ships with built-in jobs on different schedules: frequent health and commitment checks, periodic reflection, daily reviews, and infrastructure maintenance. Some jobs use an AI session; script runners use a companion .mjs file. Rollout gates can keep a converted job on its existing session path until its script is enabled.

JobCronModelPurpose
health-check*/5 * * * * (every 5 min)HaikuVerify infrastructure health
commitment-detection*/5 * * * * (every 5 min)HaikuDetect new commitments in the recent conversation flow
promise-review23 */6 * * * (every 6h at minute 23)SonnetReview up to 30 due promises, retain or prioritize uncertain work, and propose evidence-backed closures through the server. Gated by commitments.promiseReview; closures use shadow mode until graduation. Never sends messages.
commitment-checkin-reminder*/5 * * * * (every 5 min)— (no LLM)Post one reminder for each commitment whose check-in date has arrived. Ships disabled.
promise-review23 */6 * * * (every 6h, at :23)SonnetReview up to 30 due open promises; keep them, adjust priority, or propose evidence-backed completion/archive. Never sends messages. Gated by commitments.promiseReview (live on development agents, dark on the fleet); closures stay in shadow until graduation, and the server enforces evidence and per-run/day closure caps.
reflection-trigger0 */4 * * * (every 4h)OpusReflect on recent work and surface insights
evolution-overdue-check0 */4 * * * (every 4h)HaikuSurface overdue commitments and stalled action items
evolution-proposal-evaluate0 */6 * * * (every 6h)SonnetEvaluate evolution proposals against current goals
evolution-proposal-implement0 1,7,13,19 * * * (4× daily)OpusImplement evolution proposals that passed evaluation
overseer-guardian0 */6 * * * (every 6h)SonnetGuardian oversight — safety, alignment, value drift
insight-harvest0 */8 * * * (every 8h)OpusSynthesize learnings into evolution proposals
overseer-infrastructure0 6 * * * (daily 6am)HaikuInfrastructure oversight — quotas, health, scheduled work
overseer-development0 8 * * * (daily 8am)HaikuDevelopment oversight — open work, blockers, follow-through
relationship-maintenance0 9 * * * (daily 9am)HaikuReview stale relationships, refresh significance scoring
overseer-maintenance0 2 * * * (daily 2am)SonnetMaintenance oversight — log rotation, cleanup, hygiene
identity-review0 3 * * * (daily 3am)OpusIdentity review — AGENT.md drift, value alignment
overseer-learning0 3 */2 * * (every other day 3am)SonnetLearning oversight — knowledge consolidation, gap detection
jev-completion-audit0 */6 * * * (every 6h)HaikuBatch pass of the observe-only Jev job-completion audit — calls POST /jev-audit/batch, which judges captured job-completion evidence packs and writes verdict rows that decide nothing. Ships enabled: false; the endpoint is additionally inert unless intelligence.jevJobCompletionAudit is enabled inside a soak window. Excluded from its own audit.
docs-coverage-audit0 10 * * 1 (Mondays 10am)HaikuWeekly walk of the instar source tree against the docs surface, surfaces newly-undocumented capabilities. Ships enabled: false by default; only useful on machines with the instar source repo locally
org-intent-drift-audit(configurable)SonnetPeriodic drift detection for organizational intent — compares recent decisions and outputs against the constraints and goals declared in ORG-INTENT.md, surfaces drift via the degradation channel
mentor-onboarding*/15 * * * * (every 15m)HaikuFramework-Onboarding Mentor heartbeat — a thin timer that pokes POST /mentor/tick; the in-process tick runs a leak-detector canary, a fail-closed budget gate, a safe-window check, a constrained Stage-A spawn, the leakage detector, Stage-B forensics, and ledger capture. Ships enabled: false and mentor.mode: 'off'; dormant until promoted off → dry-run → live
failure-analyzer0 9 * * 3 (Wednesdays 9am)HaikuFailure-Learning Loop analyzer — weekly scan of the failure ledger for dev-process patterns. Surfaces support-and-diversity-thresholded insights, opens human-approved tracked improvements (never auto-implements), and runs the verify step on past fixes. Tier-1 supervised (wraps the deterministic /failures/analyze endpoint, validating each insight against its evidence). Ships enabled: false; turns on with monitoring.failureLearning.enabled
initiative-digest-review0 11 * * 1,4 (Mon & Thu 11am)SonnetThe self-driving half of the InitiativeTracker — twice-weekly review of the initiative board. Surfaces initiatives that need a decision and, for ships-staged features in rollout (dry-run → live → default-on), gathers promotion evidence and posts an explicit, evidence-gated recommendation. Near-silent (posts only when a genuinely-new decision is waiting); operator-gated — it recommends, it never flips a config flag
work-board-graduation20 9 * * * (daily 9:20am)HaikuWrites the Work Board pilot’s day-7 and day-14 measurement records. Its gate passes only when a record is due (twice in the pilot’s life), so on every other day, and wherever the board is off, no session starts. The server measures from the card records and writes the record; the job fires POST /work-board/graduation/pass and reports which days were written. It never messages anyone and never judges the result.
review-canary-battery15 6 * * * (daily 6:15am)HaikuDaily adversarial canary battery for the context-aware outbound review soak — drives one ReviewCanaryBattery run via the Bearer-gated POST /review/canary-battery/run trigger (seed booby-trapped fixtures into reserved negative topic ids → replay baseline + with-context arms → reviewer-level assertions → cleanup), proving the “user asked for this” carve-out cannot launder a credential/PII paste past the opted-in reviewer. Every outcome — including refusals — writes a batterySummary row, so a silent skip is impossible. Tier-1 supervised. Ships enabled: false; the operator enables it only on the soaking dev agent for the §D9 soak window
bench-refresh0 4 1 * * (monthly, 1st 4am)HaikuLLM routing benchmark refresh — on the machine carrying the bench harness it reruns the benchmark + a parity-check and raises ONE operator-review diff when a routing default looks stale. Never auto-applies a routing change; no-ops on any machine without the harness. Tier-1 supervised. Ships enabled: false; enable only on a maintainer machine
doorway-scan0 4 * * 1 (Mondays 4am)HaikuDoorway/Model Knowledge Registry live re-probe — a deterministic prober (scripts/doorway-scan.mjs) probes each doorway (which CLIs / free model-list APIs answer, and which top models each exposes), updates the per-machine live scan-state, diffs against the previous scan, and raises ONE jargon-safe operator diff with only the changes. Free probes by default (zero metered spend); never auto-edits the canonical registry. perMachineIndependent (each machine scans its own disk). Tier-1 supervised. Ships enabled: false; enable only on a maintainer machine
routing-price-refresh0 5 * * 1 (Mondays 5am)HaikuRouting Control Room spend view (Increment A) price refresh — a deterministic prober (scripts/routing-price-refresh.mjs) re-confirms published per-token prices for the metered routing doors from PUBLIC, no-auth model-list endpoints (OpenRouter) at the free scope, validates each price (range + cached≤input + UTC day-alignment), and writes forward-only points into the MACHINE-LOCAL observed cache (.instar/routing-prices.observed.json) ONLY — structurally never the canonical price manifest. Observed points feed the read-only spend view + a promote-me drift hint; they are gate-ineligible by construction. Metered/web-verify probes are manual-only + budget-fail-closed. perMachineIndependent. Tier-1 supervised. Ships enabled: false; enable only on a maintainer machine
routing-price-web-verify0 6 * * 2 (Tuesdays 6am)HaikuThe SCHEDULED web-research price check (operator directive 2026-07-07) for the metered doors whose prices live on OFFICIAL WEB PAGES only (Groq, Google): the same deterministic prober at --scope +web-verify fetches groq.com/pricing + ai.google.dev/pricing, extracts tracked-model prices with conservative fail-closed parsers (fixture-realness-tested — a reshaped page refuses, never guesses) plus a >10x plausibility clamp vs the reviewed canonical price, and writes forward-only OBSERVATIONS into the machine-local observed cache only. Zero spend (no LLM, no metered key). An observed price never becomes official without the operator’s PIN promotion. perMachineIndependent. Tier-1 supervised. Ships enabled: false
benchmark-divergence-analysis45 3 * * * (daily 3:45am)— (tier-0, deterministic)Benchmark-Divergence Detector analysis pass — triggers POST /benchmark-divergence/analyze with {"trigger":"cadence"}. The server-side pass is fully deterministic (rolling matured-day window over the per-model quality rollup, pool-collect across machines, Wilson-aware comparison against the mirrored benchmark predictions, idempotent advisory finding upserts), so tier-0 supervision is justified — there is no LLM step anywhere. Lease-gated (a non-holder answers 409 and the job exits silently); observe-only (never messages). Ships enabled: false fleet-wide; the rollout step enables it on the development agent for the dryRun soak
feedback-factory-triage*/15 * * * * (every 15 min)HaikuFeedback triage tick — calls POST /feedback-factory/triage/tick so the operator-approved triage authority decides work / hold / ignore for feedback work items within deterministic floors, and held items come back on their timers. Near-silent; tier-1 supervised. Development-agent live, fleet dark; inert until the operator approves the triage authority. See feedback factory processing
feedback-factory-action-list0 8 * * * (daily 8am)HaikuOnce-a-day operator action list from feedback triage — calls POST /feedback-factory/triage/action-list, which sends only NEW items that need the operator (at most 10 plus a count line, each with a direct link). The server enforces the 23:00–07:30 quiet window and once-per-item delivery. See feedback factory processing
feedback-factory-execute17,47 * * * * (twice hourly)Gated scriptReconciles feedback fix attempts through POST /feedback-factory/execute/tick. The current script rollout keeps its prior session path until this slug is enabled; server execution starts in development dry-run, and the fleet route remains dark.
script-job-supervisor23 7 * * * (daily)Script + bounded AI reviewReviews script-job decisions and missing-machine signals once daily when the script rollout is active. It stays inactive on agents where the rollout is off.

All jobs ship inside src/scaffold/templates/jobs/instar/ and are installed on instar init plus refreshed on every update via PostUpdateMigrator.

Jobs declare a supervision field that controls how each step is validated:

  • tier0 — Raw programmatic. No LLM validation. Fast, cheap, silent failures.
  • tier1 — LLM-supervised. A lightweight model (typically Haiku) validates each step. Observed failures.
  • tier2 — Full intelligent. A capable model (Sonnet or Opus) handles reasoning end-to-end. Handled failures.

The supervision tier is independent of the execution model — tier1 may use Haiku while the job runs on Sonnet, for instance. See docs/LLM-SUPERVISED-EXECUTION.md for the full design.

The scheduler reads from a shared QuotaTracker and shedds load as quota tightens. Threshold buckets in scheduler.quotaThresholds:

  • normal — full scheduling
  • elevated — defer Opus-tier jobs
  • critical — defer Sonnet-tier jobs as well
  • shutdown — pause everything except health-check

Tier-aware shedding lets you keep critical safety jobs alive even when you’re hammering against your daily cap.

When the host wakes from sleep, the scheduler reaps any pending runs older than wakeReaper.thresholdMultiplier × expectedDurationMinutes. This prevents a stampede of overdue jobs from firing all at once after a long suspend.

Job gates (preconditions evaluated before the job body runs) can fail transiently. The scheduler retries gates up to gateRetries times (default 3) with gateRetryDelayMs between attempts (default 5 s). Persistent gate failures surface as a degradation rather than a stuck job.

Edit .instar/jobs.json (or the per-job .md files under .instar/jobs/instar/ if your agent uses the newer agentmd execution type) to:

  • Change schedules
  • Adjust models
  • Add new jobs
  • Disable jobs you don’t need

The agent can also modify its own jobs through the evolution system.

A newer execution type, agentmd, lets the job body live in the markdown frontmatter of .instar/jobs/<origin>/<slug>.md rather than inline in jobs.json. Built-in jobs ship as Markdown files under src/scaffold/templates/jobs/instar/. A file can also declare runner: script and name a companion script installed under .instar/jobs/instar/scripts/.

These older jobs still exist in some agent installations for backward compatibility but are disabled by default:

JobReplaced by
update-checkAutoUpdater (built-in server component — no session needed)
dispatch-checkAutoDispatcher (built-in server component — no session needed)

Both replacements are server-side components that don’t spawn Claude sessions, saving quota.

The failure-callout-investigator job investigates serious agent failures using the correction ledger. It runs at minutes 11 and 41, reserves a bounded batch before starting a model session, and retains the reservation during capacity queuing. Three unchanged attempts exhaust the automatic budget; a timer cannot reset it. The agent must provide diagnosis, class repair and current verification before closure. New occurrences reopen the case. Read the correction learning guide for evidence, origin ownership and opt-out behavior.

The separate promise-review job reviews tracked commitments against their actual state and records progress or evidence through the commitment system. It is not a substitute for investigating the root cause of a failure and does not authorize closing a failure case without its replay and legitimate-behavior checks. Both jobs keep unfinished work visible without asking the user to manage a second task list.