Minion Orchestrator

Contract

Minions is a Postgres-native job queue for durable, observable background work. This single skill handles two lanes:

• Deterministic shell jobs (gbrain jobs submit shell ...)
• LLM subagent jobs (gbrain agent run ...)

When to route to Minions: durable, observable work that must survive restarts, fan out across many parallel tasks, or persist across sessions. Routing policy is defined in skills/conventions/subagent-routing.md — the project default is pain_triggered (native subagents first, Minions after specific pain signals fire); Mode A (all-through-Minions) is opt-in.

Guarantees:

• Jobs survive gateway restart (Postgres-backed)
• Every job has structured progress, token accounting, and session transcripts
• Running agents can be steered mid-flight via inbox messages
• Jobs can be paused, resumed, or cancelled at any time
• Parent-child DAGs with configurable failure policies

Durable-execution doctrine (routing convention the agent follows — nothing mechanically enforces it; see "Durable execution" below):

• Operations expected to exceed ~2 minutes route through the capability ladder, never a bare background shell that dies with the session.
• A deadman check verifies the result was REPORTED to the user, not merely that the process exited — a swallowed completion event looks identical to a dead job from the user's side.
• The deadman is second-line insurance, not a delivery guarantee: it can itself die before firing (scheduler restart, deleted entry). Backstop: on the next turn, sweep gbrain jobs list --status active and recent completions for overdue work whose result never reached the user.
• Deadman checks are idempotent. A double-fire, or a fire that races the normal completion report, produces silence — it checks reported-state first and tags any recovery report with the job ID so duplicates are detectable.
• A checkpoint or progress file is trusted only with a freshness check. A stale checkpoint reads as "still running" forever; compare its timestamp against the expected progress interval and against gbrain jobs get <id>.

Route the Request: Shell Job vs Subagent

If intent is ambiguous, ask one clarification: "Do you want a deterministic shell command job, or an LLM agent job?"

Shell Jobs (Deterministic Scripts)

Use for reproducible command execution, ETL steps, cron work, and scriptable tasks where no LLM reasoning loop is needed.

Preconditions (read before submitting your first shell job)

• GBRAIN_ALLOW_SHELL_JOBS=1 must be set on the worker environment. Without it, the shell handler refuses to register and submissions sit in waiting silently. Gate lives in src/core/minions/handlers/shell.ts.
• Security: flipping GBRAIN_ALLOW_SHELL_JOBS=1 authorizes arbitrary command execution on the worker. On a shared queue, this is a remote code execution surface. Treat as privileged infrastructure authorization.
• Execution mode — pick one:
  • Postgres + daemon: gbrain jobs work runs a persistent worker that claims and executes jobs from the queue.
  • PGLite + --follow: gbrain jobs submit ... --follow runs inline. The daemon mode is not available on PGLite (exclusive file lock). See docs/guides/minions-shell-jobs.md.
• MCP boundary: shell-job submission is CLI-only. submit_job name="shell" over MCP throws an OperationError with code permission_denied ("'shell' jobs cannot be submitted over MCP") because shell is in PROTECTED_JOB_NAMES. Agents CAN observe shell jobs via get_job / list_jobs / get_job_progress (not protected), but cannot submit them. Operator or autopilot submits; agent observes.
• Verify setup: after configuration, run gbrain jobs stats (CLI) to confirm the worker is registered and consuming the queue.

Submit (CLI, operator or autopilot)

Shell jobs take their command via --params as a JSON object with cmd (string) or argv (array), plus cwd and optional env.

Command string form:

gbrain jobs submit shell --params '{"cmd":"echo hello","cwd":"/abs/path"}'

Argv form (no shell expansion):

gbrain jobs submit shell --params '{"argv":["bash","-lc","echo hello"],"cwd":"/abs/path"}'

Inline execution on PGLite or any one-shot deployment:

gbrain jobs submit shell --params '{"cmd":"echo hello","cwd":"/tmp"}' --follow

Queue/lifecycle flags exposed by gbrain jobs submit --help: --queue, --priority, --delay, --max-attempts, --max-stalled, --backoff-type, --backoff-delay, --backoff-jitter, --timeout-ms, --idempotency-key, --dry-run.

Monitor (agents or operator)

These operations are MCP-callable and safe for agent use:

list_jobs --name shell --status active
get_job ID
get_job_progress ID

Check structured result fields (exit code, stdout/stderr tails, attempts, timings) from get_job. Use get_job_stats (MCP) or gbrain jobs stats (CLI) for the worker/queue health dashboard incl. the wedged-queue signal.

Control (MCP-callable)

cancel_job id=ID
replay_job id=ID

replay_job is not protected — only shell submission is. Agents can cancel or replay a shell job without CLI access.

Use idempotency keys for recurring shell workloads to avoid duplicate runs.

Subagent Jobs (LLM Orchestration)

Use for open-ended reasoning, tool-using research, and fan-out synthesis.

User-facing entrypoint: gbrain agent run <prompt> is the canonical way to submit subagent work. It handles the elevated-trust plumbing — subagent and subagent_aggregator are both in PROTECTED_JOB_NAMES, so direct MCP submission requires {allowProtectedSubmit: true}, which gbrain agent run supplies.

Phase 1: Submit

gbrain agent run "Research Acme Corp revenue" --tools "search,query"

--tools accepts a comma-separated subset of BRAIN_TOOL_ALLOWLIST (see src/core/minions/tools/brain-allowlist.ts): query, search, get_page, list_pages, file_list, file_url, get_backlinks, traverse_graph, resolve_slugs, get_ingest_log, put_page. Anything outside the allow-list is rejected at submit time with allowed_tools references unknown tool.

For parallel work with a fan-out manifest:

gbrain agent run --fanout-manifest companies.json

The manifest describes N children + 1 aggregator. Each child runs name="subagent" under the hood; the aggregator runs name="subagent_aggregator" and claims AFTER every child terminates. See src/core/minions/handlers/subagent.ts and src/core/minions/handlers/subagent-aggregator.ts.

Flags (from src/commands/agent.ts):

• --subagent-def <name> — named subagent definition
• --model <id> — override model
• --max-turns <N> — cap the LLM loop
• --tools <csv> — allow-listed brain tools (see above)
• --timeout-ms <N> — hard timeout per job
• --fanout-manifest <file> — N children + 1 aggregator
• --follow / --no-follow — stream logs + wait (default on TTY)
• --detach — submit and return immediately

Queue/priority/retry tuning is not exposed by gbrain agent run; submit the raw subagent handler via gbrain jobs submit (requires CLI trust) if you need those knobs.

Admission control (v0.46.11.0). Identical parentless subagent submits (same owner lane, payload, and execution options) coalesce onto the existing waiting job: gbrain agent run prints coalesced with the matched job id, and the submit_agent MCP response carries coalesced: true. Treat that as success — monitor the matched id, do NOT resubmit. Jobs still waiting after the TTL (48h default for subagent; minions.ttl_waiting_hours.<name>) are cancelled with reason prefix waiting_ttl_expired. If an operator has configured a waiting quota (minions.quota_max_waiting.<name>), a submit past the cap returns a structured, retryable rate_limited error — back off and check gbrain jobs stats for a DIVERGENT QUEUE line before retrying.

Phase 2: Monitor

list_jobs --status active          # MCP — what's running?
get_job ID                         # MCP — full details + logs + tokens
get_job_progress ID                # MCP — structured progress snapshot
gbrain jobs stats                  # CLI — queue health dashboard
gbrain agent logs ID --follow      # CLI — streaming transcript + heartbeat

Progress includes: step count, total steps, message, token usage, last tool called.

Phase 3: Steer

Send a message to redirect a running agent:

send_job_message id=ID payload={"directive":"focus on revenue, skip headcount"}

The agent handler reads inbox messages on each iteration and injects them as context. Messages are acknowledged (read receipts tracked).

Only the parent job or admin can send messages (sender validation).

Phase 4: Lifecycle

pause_job id=ID                    # freeze without losing state
resume_job id=ID                   # pick up where it left off
cancel_job id=ID                   # hard stop
replay_job id=ID                   # re-run with same or modified params
replay_job id=ID data_overrides={"depth":"deep"}  # replay with changes

All lifecycle ops are MCP-callable.

Phase 5: Review Results

get_job ID                         # result, token counts, transcript

Token accounting: every job tracks tokens_input, tokens_output, tokens_cache_read. Child tokens roll up to parent automatically on completion.

Durable execution (operations >2 minutes)

Background shells die silently: session compaction, harness restart, tool timeout. Long gbrain operations (extract all, embed --stale, a full sync --all) routinely run 10-60 minutes — past every one of those ceilings. And even when the work survives, the completion event can be swallowed (worker restart, dropped notification), leaving the user staring at silence while a finished result sits unreported. Durable execution covers both halves: the work survives, and the report provably lands.

Route any operation expected to exceed ~2 minutes through the highest rung of this ladder the deployment supports. This is a harness-routing convention the agent follows, not a mechanical guarantee — nothing stops a bare background shell except this skill saying don't.

Rung 1 — Minion job + deadman (Postgres + worker)

Requires: Postgres engine, a running gbrain jobs work worker, and — for the shell lane — GBRAIN_ALLOW_SHELL_JOBS=1 on the worker. All the Preconditions above still hold: the flag defaults OFF, shell submission is CLI-only across the MCP trust boundary, and PGLite has no worker daemon (see Rung 3). Nothing in this section loosens that contract.

1. Submit as a job so the work survives restarts:

   gbrain jobs submit shell \
     --timeout-ms 3600000 \
     --params '{"cmd":"gbrain extract all","cwd":"/abs/path","inherit":["database_url"]}'

   Work that needs LLM judgment goes through the subagent lane (gbrain agent run, above) instead — a submitted agent job you never check on is the same bug as an unwatched shell job.

2. Arm a deadman in the same action block (pattern below). Submitting without arming is the classic half-fix: the job survives, the silence doesn't.

Rung 2 — cron-checked progress file (no agent-waking scheduler)

When no scheduler can wake the agent but the host has a plain crontab (or the work runs outside the queue entirely): make the operation write a progress/heartbeat file as it advances (or rely on get_job_progress for queue jobs), and register a recurring host-cron check that compares the file's freshness against the expected progress interval. The agent also checks it at the start of the next turn. A checkpoint that stops advancing means stalled, not "still running" — the freshness comparison is the entire value of this rung.

Rung 3 — foreground + output buffering (PGLite / no worker / no scheduler)

Run the operation inline in the foreground — on PGLite that's gbrain jobs submit ... --follow, or just the raw command — and buffer all output to a file, reading bounded slices, per skills/conventions/exec-output.md. An empty tool result after a long command is truncation, not a dead shell. This rung has no silent-death insurance, so keep the operation in the foreground and stay with it; backgrounding here recreates the exact failure the ladder exists to prevent.

The deadman pattern

A one-shot, self-deleting scheduled check that fires at expected-finish-plus-margin and verifies the result was reported — not just that the process exited. Arm it in the same action block as the submission (not after, not "if I remember").

1. Estimate expected_minutes honestly; round up.
2. Deadline = now + estimate + margin, where margin = max(10 min, 50% of estimate). Restarts delay delivery — too tight false-fires, hours-late defeats the point.
3. Create the check with whatever one-shot scheduler the host offers: a harness cron tool with delete-after-run semantics, at, or a crontab entry the check removes on first fire. The check's instruction:
   • Verify the result was already reported to the user. gbrain jobs get <id> gives the job state; the reported-check asks whether a completion message actually reached the user.
   • Reported → do nothing, stay silent, self-delete.
   • Completed but unreported (swallowed event) → pull the result via gbrain jobs get <id> and post the recovery report now, tagged with the job ID.
   • Still running → post a one-line status with a new ETA and arm ONE follow-up deadman at +50% of the original estimate. One follow-up max — no infinite chains.
   • Dead/failed → report what it produced before dying (stderr_tail from gbrain jobs get <id>) and offer gbrain jobs retry <id>.
4. Disarm on normal completion. When the completion arrives and you report it, remove the deadman. If it fires anyway, the reported-check makes it a silent no-op — belt and suspenders.

Failure modes the pattern must own (mirrored in Contract and Anti-Patterns): the deadman itself dying before it fires (second-line insurance, backstopped by the next-turn overdue sweep), double-fire (idempotent reported-check + job-ID-tagged reports), and stale checkpoints (freshness check before trusting "still running").

Eval contract, imported with the pattern — a deadman deployment is judged on:

• ARMED_ON_SPAWN — created in the same action block as the long submission?
• SELF_DELETING — one-shot, and silent when all is well?
• DETECTS_SWALLOWED — verifies the result was reported, not just that the job exited?
• RECOVERS — on a swallowed event, pulls the output and posts the recovery report?
• RIGHT_DEADLINE — expected finish + sane margin, neither false-firing nor hours late?

Hard fails: a long user-facing operation with no deadman armed; a deadman that posts noise when the completion arrived normally; emulating the timer with sleep or a poll loop instead of a scheduler entry (a sleeping process dies with the session — the exact failure being insured against).

Sequencing and locks

Maintenance operations share locks (sync, embed, extract, integrity). Run them sequentially, chained: submit job 1, arm its deadman; on completion, submit job 2, arm the next; finish with gbrain doctor and report the health delta. If a job dies mid-operation its lock expires at TTL (a live, recently-refreshed holder is protected by the steal grace) — never hand-delete lock rows to "unstick" a queue.

Typical long operations

Durations scale with corpus size; treat these as order-of-magnitude anchors for --timeout-ms, not promises.

Appendix: content-addressed stage checkpoints

For a multi-stage pipeline with an expensive middle (extract → score → explain → render → verify, where the scoring stage burns real LLM spend), make each stage a content-addressed checkpoint so a crash — or a deadman-triggered retry, or a replay_job — resumes instead of re-spending:

• Key each stage on a hash of (the stage's own logic + its params + the hashes of its upstream artifacts). Store artifacts under a .cache/ directory in the pipeline's working tree, addressed by content hash.
• Warm re-run is a no-op: every key matches, nothing recomputes, the whole pipeline replays in well under a second.
• Busting cascades correctly: a param change or a logic edit changes that stage's key, recomputing it and every downstream stage whose upstream hashes changed. The stage's own source must be part of the key — omit it and logic edits silently reuse stale artifacts.
• Record provenance: a provenance file per run records each stage's key, inputs, outputs, and computed-at, so every artifact traces to the exact logic + data that produced it and "which stage recomputed and why" stays answerable.
• Verify integrity on restore: check live artifacts against recorded checksums; a mismatch means recompute or restore from cache, never silent reuse.

Judged on: IDEMPOTENT (warm re-run recomputes nothing), CORRECT_BUSTING (a change recomputes exactly the affected stages), PROVENANCE (every artifact traces to logic + params + upstreams), INTEGRITY (corrupted artifacts detected, never silently reused).

This composes with the ladder rather than replacing it: the ladder keeps the pipeline running and reported; stage checkpoints keep a retry cheap. Pair them whenever a single stage costs more than pocket change in LLM spend.

Output Format

When reporting job status to the user:

Job #ID (name) — status
Progress: step/total — last action
Tokens: input_count in / output_count out (+ cache_read cached)
Runtime: Xs
Children: N pending, M completed

When reporting completion:

Job #ID completed in Xs
Tokens used: input / output / cache_read
Result: <summary>

When reporting batch status (parent with children):

Parent #ID — waiting-children
  #A subagent(Acme) — active, 3/5 steps, 2.5k tokens
  #B subagent(Beta) — completed, 1.8k tokens
  #C subagent(Gamma) — paused
Total tokens so far: 4.3k

Anti-Patterns

• Don't spawn a Minion for a single search query (use search tool directly)
• Don't fire-and-forget without checking results
• Don't spawn > 5 concurrent agents without checking gbrain jobs stats first
• Don't resubmit when a submit reports coalesced — the work is already queued; monitor the matched job id instead
• For subagent work, don't use sessions_spawn with runtime: "subagent" when Minions is available (use gbrain agent run instead)
• Don't poll get_job in a tight loop (use get_job_progress for lightweight checks)
• Don't run an operation expected to exceed ~2 minutes as a bare background shell — it dies with the session; route through the Durable execution ladder
• Don't say "I'll report back when it finishes" without an armed deadman or a scheduled check that will actually fire
• Don't let a deadman post noise when the completion arrived normally — check reported-state first, stay silent, self-delete
• Don't emulate the deadman timer with sleep or a poll loop — a sleeping process dies with the session, which is the exact failure being insured against
• Don't trust a checkpoint or progress file without a freshness check — a stale checkpoint reads as "still running" forever
• Don't run lock-acquiring maintenance ops (sync, embed, extract, integrity) simultaneously, and don't hand-delete lock rows to unstick them (locks expire at TTL)

Tools Used

• Submit a background job — submit_job (MCP, non-protected names only; shell jobs are CLI-only, subagent jobs via gbrain agent run)
• Get job details — get_job (MCP)
• List jobs with filters — list_jobs (MCP)
• Cancel a job — cancel_job (MCP)
• Pause a job — pause_job (MCP)
• Resume a paused job — resume_job (MCP)
• Replay a completed/failed job — replay_job (MCP)
• Send sidechannel message — send_job_message (MCP)
• Get structured progress — get_job_progress (MCP)
• Queue stats — get_job_stats (MCP; admin scope over HTTP, same as the other jobs ops here — includes the wedged-queue silent-halt signal) or gbrain jobs stats (CLI)