Default Jobs
Instar ships with 26 default jobs that give your agent a circadian rhythm — regular self-maintenance, monitoring, evolution, and growth without user intervention. Every job was born from a real need: sessions that stalled unnoticed, learnings that were never captured, state that drifted between machines.
Jobs are organized into six categories. Each category has worker jobs that do the actual work, and most have an overseer that reviews the category’s output and spots cross-job patterns.
At a Glance
Section titled “At a Glance”| Category | Jobs | What They Do |
|---|---|---|
| Guardian | 7 | Watch the agent’s health and catch problems |
| Learning | 7 | Grow smarter through reflection, evolution, and commitment tracking |
| Maintenance | 5 | Keep state files clean and maps current |
| Infrastructure | 3 | Sync data, refresh links, retry failures |
| Identity | 1 | Review identity coherence and soul drift |
| Overseers | 5 | Supervise each category — the jobs that watch the jobs |
Guardian & Monitoring
Section titled “Guardian & Monitoring”These jobs detect problems before users do. They work alongside the system processes that run continuously — the processes detect in real-time, the jobs review patterns over time.
health-check
Section titled “health-check”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 5 min | Haiku | Critical | ~1 min |
The heartbeat. Verifies the instar server is responding, checks disk space, and reports issues. Silence means healthy — it only messages you when something needs attention.
Works with: HealthChecker (continuous process) and CoherenceMonitor (5-minute integrity sweeps).
guardian-pulse
Section titled “guardian-pulse”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 8 hours | Haiku | High | ~2 min |
The meta-monitor. Verifies that other jobs are running, healthy, and not silently failing. If health-check itself stops running, guardian-pulse is the one that notices.
Works with: The job scheduler execution ledger — reads run history to detect skips, failures, and scheduling drift.
degradation-digest
Section titled “degradation-digest”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 4 hours | Haiku | Medium | ~1 min |
Groups DegradationReporter events into patterns. A single fallback is noise; the same fallback happening every hour is a trend worth escalating.
Gate: Only runs when degradation events exist.
state-integrity-check
Section titled “state-integrity-check”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Haiku | Medium | ~1 min |
Cross-validates state file consistency. Detects orphaned references (a session ID in one file that doesn’t exist in another), bloated files, and structural corruption.
session-continuity-check
Section titled “session-continuity-check”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 4 hours | Haiku | Medium | ~2 min |
Verifies that sessions produce lasting artifacts — handoff notes, memory updates, learnings. A session that runs for 30 minutes and leaves no trace is a wasted session.
Works with: SessionActivitySentinel (generates the activity digests this job reviews).
commitment-detection
Section titled “commitment-detection”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 5 min | Haiku | High | ~1 min |
Scans recent Telegram messages for promises and commitments (“I’ll build that tomorrow”, “let me deploy on Friday”). Registers them as evolution actions so they can be tracked. Replaced the server-side CommitmentSentinel process — same intelligence, but as a scheduled job that’s easier to tune.
Works with: evolution-overdue-check (the paired job that monitors whether commitments are kept).
overseer-guardian
Section titled “overseer-guardian”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Sonnet | Medium | ~5 min |
Reviews all guardian jobs as a group. Spots cross-job patterns: Is health-check reporting healthy while degradation-digest finds issues? Are jobs being skipped excessively? Could any use a cheaper model? Only alerts on findings — silence means the monitoring system is healthy.
Learning & Evolution
Section titled “Learning & Evolution”The growth engine. These jobs implement a pipeline: detect learnings → synthesize insights → propose improvements → evaluate → implement → track commitments.
reflection-trigger
Section titled “reflection-trigger”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 4 hours | Opus | Medium | ~5 min |
Reviews recent activity and updates MEMORY.md with any learnings worth preserving. The raw material for the learning pipeline — everything downstream depends on high-quality reflections.
insight-harvest
Section titled “insight-harvest”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 8 hours | Opus | Low | ~3 min |
Synthesizes unapplied learnings into patterns. Multiple learnings pointing to the same conclusion become an evolution proposal. Cross-domain connections (insight from messaging applied to scheduling) are the highest-value finds.
Works with: EvolutionManager (the server-side system that manages the proposal lifecycle).
evolution-proposal-evaluate
Section titled “evolution-proposal-evaluate”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Sonnet | Medium | ~3 min |
Phase A of the evolution pipeline. Reads pending proposals, evaluates merit, and approves or rejects. Does NOT implement — that’s the paired job’s responsibility.
Gate: Only runs when proposals with status proposed exist.
evolution-proposal-implement
Section titled “evolution-proposal-implement”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| 4x daily (midnight, 6am, noon, 6pm) | Opus | Medium | ~10 min |
Phase B of the evolution pipeline. Picks up approved proposals and implements them — creates skills, hooks, job modifications, or config changes. The most expensive job by design: implementation requires deep context.
Gate: Only runs when proposals with status approved exist.
evolution-overdue-check
Section titled “evolution-overdue-check”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 4 hours | Haiku | High | ~2 min |
Monitors overdue evolution actions and stale commitments. If the agent promised to do something by Friday and it’s Saturday, this job notices. Also catches items that have been pending for 48+ hours without a due date — forgotten commitments.
Paired with: commitment-detection (detects promises) and EvolutionManager (tracks action lifecycle).
identity-review
Section titled “identity-review”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 3 AM | Opus | Medium | ~5 min |
Reviews identity coherence. Checks whether AGENT.md and soul.md have drifted from the agent’s actual behavior, and nudges reflection if identity-relevant learnings have accumulated.
overseer-learning
Section titled “overseer-learning”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 2 days at 3 AM | Sonnet | Medium | ~5 min |
Reviews the entire learning pipeline. Key question: is the agent actually getting smarter over time, or is the learning pipeline just busy-work? Tracks proposal accept/reject ratios, insight novelty, commitment completion rates, and memory update quality.
Maintenance
Section titled “Maintenance”Housekeeping jobs that keep state files clean and maps current. Good maintenance trends toward finding less work over time — if these jobs consistently find issues, something upstream is broken.
project-map-refresh
Section titled “project-map-refresh”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 12 hours | Haiku | Medium | ~1 min |
Regenerates the project territory map — the agent’s spatial awareness of its working directory. Script-type job (no LLM session needed).
coherence-audit
Section titled “coherence-audit”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 8 hours | Haiku | Medium | ~2 min |
Verifies topic-project bindings are valid, state files are healthy, and no drift has occurred.
Works with: CoherenceMonitor (continuous 5-minute checks). The process checks structural integrity; this job checks semantic coherence.
capability-audit
Section titled “capability-audit”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Haiku | Medium | ~1 min |
Refreshes the capability map and detects drift. Compute-first: compares the manifest to current state and only spawns an LLM if changes are detected. Script-type job.
memory-hygiene
Section titled “memory-hygiene”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 12 hours | Opus | High | ~5 min |
Reviews MEMORY.md for stale entries, duplicates, contradictions, and quality issues. Proposes cleanup. Uses Opus because memory quality directly impacts every other job that reads MEMORY.md.
Gate: Only runs when MEMORY.md exceeds 100 words.
memory-export
Section titled “memory-export”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Haiku | Medium | ~1 min |
Regenerates MEMORY.md from the SemanticMemory knowledge graph. Keeps the human-readable snapshot fresh without manual intervention. Script-type job.
Works with: MemoryMonitor (tracks memory system health).
overseer-maintenance
Section titled “overseer-maintenance”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 2 AM | Sonnet | Medium | ~5 min |
Reviews all maintenance jobs. Watches for diminishing returns — if memory-hygiene consistently finds nothing, it should run less often. Recommends frequency adjustments and flags redundant checks.
Infrastructure
Section titled “Infrastructure”The plumbing. These jobs should be boring and reliable — any excitement is a problem.
git-sync
Section titled “git-sync”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every hour | Haiku | High | ~5 min |
Intelligent multi-machine git synchronization. Pulls remote changes, merges with conflict resolution, commits local changes, and pushes. Uses tiered model escalation: Haiku for clean syncs, Sonnet for state file conflicts, Opus for code conflicts.
Works with: GitSyncManager (the server-side manager that handles debounced commits and sync cycles). The process manages real-time syncing; this job ensures periodic full reconciliation.
dashboard-link-refresh
Section titled “dashboard-link-refresh”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 15 min | Haiku | Medium | ~1 min |
Refreshes the pinned dashboard link in Telegram so it never goes stale (tunnel URLs can change). Script-type job — no LLM needed.
feedback-retry
Section titled “feedback-retry”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Every 6 hours | Haiku | Low | ~1 min |
Retries forwarding any feedback items that failed to reach the upstream API. Script-type job.
Works with: DeliveryRetryManager (handles message-level retries). This job handles feedback-specific retries.
overseer-infrastructure
Section titled “overseer-infrastructure”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 6 AM | Haiku | Medium | ~3 min |
Reviews infrastructure jobs. Checks git-sync success rates, dashboard link freshness, and whether feedback-retry is actually retrying anything. The quietest overseer — infrastructure should just work.
Identity
Section titled “Identity”identity-review
Section titled “identity-review”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 3 AM | Opus | Medium | ~5 min |
Reviews identity coherence — checks whether the agent’s behavior aligns with its AGENT.md and soul.md files. Nudges reflection if identity-relevant learnings have accumulated. Uses Opus because identity questions require nuanced judgment.
Overseers
Section titled “Overseers”Every category has a supervisor job that reviews the group’s output and spots patterns individual jobs can’t see.
| Overseer | Watches | Schedule | Model |
|---|---|---|---|
| overseer-guardian | health-check, guardian-pulse, degradation-digest, state-integrity-check, session-continuity-check | Every 6h | Sonnet |
| overseer-learning | reflection-trigger, insight-harvest, evolution-*, commitment-detection | Every 2 days | Sonnet |
| overseer-maintenance | project-map-refresh, coherence-audit, capability-audit, memory-hygiene, memory-export | Daily | Sonnet |
| overseer-infrastructure | git-sync, dashboard-link-refresh, feedback-retry | Daily | Haiku |
| overseer-development | ci-monitor (if present) | Daily | Haiku |
overseer-development
Section titled “overseer-development”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 8 AM | Haiku | Low | ~3 min |
Reviews development-focused jobs (like ci-monitor, if the agent has one). Only valuable when there’s active development — if the codebase is stable, this overseer stays quiet.
Relationships
Section titled “Relationships”relationship-maintenance
Section titled “relationship-maintenance”| Schedule | Model | Priority | Duration |
|---|---|---|---|
| Daily at 9 AM | Haiku | Low | ~3 min |
Reviews tracked relationships and surfaces observations about stale contacts — anyone with significance ≥ 3 who hasn’t been heard from in over 2 weeks.
Works with: RelationshipManager (core server component that stores and queries relationship data).
Model Tiering
Section titled “Model Tiering”Jobs use different models based on what they need to do:
| Model | Cost | Used For | Jobs |
|---|---|---|---|
| Haiku | Low | Quick checks, script gating, pattern matching | health-check, guardian-pulse, commitment-detection, degradation-digest, state-integrity-check, session-continuity-check, coherence-audit, capability-audit, project-map-refresh, memory-export, git-sync, dashboard-link-refresh, feedback-retry, relationship-maintenance, overseer-infrastructure, overseer-development |
| Sonnet | Medium | Evaluation, synthesis, cross-job analysis | evolution-proposal-evaluate, overseer-guardian, overseer-learning, overseer-maintenance |
| Opus | High | Deep reasoning, implementation, identity | reflection-trigger, insight-harvest, evolution-proposal-implement, memory-hygiene, identity-review |
Execution Types
Section titled “Execution Types”Not every job needs a Claude session:
| Type | What It Does | Example |
|---|---|---|
| prompt | Spawns a Claude session with the prompt as input | reflection-trigger, health-check |
| script | Runs a shell script directly — no LLM, no quota | feedback-retry, dashboard-link-refresh, capability-audit |
| skill | Invokes a Claude Code skill (slash command) | git-sync, memory-hygiene, coherence-audit |
Script-type jobs are free — they don’t consume API quota. When a job can be done without LLM reasoning (API calls, file checks, data transforms), it should be a script.
Customization
Section titled “Customization”Edit .instar/jobs.json to:
- Change schedules or models
- Adjust priorities
- Add custom jobs
- Disable jobs you don’t need
The agent can also modify its own jobs through the evolution system — proposing schedule changes, model downgrades, or entirely new jobs based on observed patterns.
See Also
Section titled “See Also”- Under the Hood — The 54 system processes that run alongside these jobs
- The Living System — How jobs and processes work together as a unified organism
- Job Scheduler — How the scheduler decides what to run and when