Alex Reysa Pm Go Pm Go Orchestration

pm-go Orchestration

pm-go pulls planning, retries, approvals, budgets, worktree lifecycle, merge order, and completion audit out of model context and into a durable control plane (Postgres + Temporal + git worktrees + typed contracts). Runs are resumable; "done" comes from evidence (diff-scope, reviewer findings, deterministic audit) rather than model claims.

Operating Defaults

Prefer the CLI and typed operator tools over raw API calls. They preserve path containment, approval policy, process cleanup, and plan provenance. Use raw API only for manual recovery or when the user is explicitly working at the API layer; the current endpoint cheatsheet lives in references/api-surface.md.

When state is unclear, inspect first with pm-go status or pm-go why <id>. When command syntax is unclear, run pm-go <command> --help rather than inventing flags.

Activation Loop (start here)

pm-go is agent-first as of v0.9.x. The default invocation is the agentic operator: it ensures Docker + Postgres + Temporal + worker + API are reachable, then drives the control plane through typed MCP tools.

pm-go implement without --legacy-drive also defaults to agent mode since v0.9. The pre-v0.9 "boot + submit + drive in one process" semantics live behind --legacy-drive, including its 45-minute default plan-persistence wait (--plan-wait 60m overrides it).

Stay in the foreground; Ctrl+C tears it down cleanly. For the default agent, --approve interactive is the default; --yes changes the default approval policy to all; an explicit --approve all|none|interactive always wins. The legacy implement and low-level drive commands default to --approve all.

Do not interrogate the user about env vars before running. Trust --runtime auto (the default). It auto-detects, in priority order:

1. ANTHROPIC_API_KEY env var
2. Claude Code OAuth session at ~/.claude/.credentials.json (i.e. the user is logged into Claude Code)
3. claude CLI on PATH

If none of those are present, the worker fails loudly at boot with an actionable error. You don't need to pre-check — the supervisor's first banner line tells you which auth source was selected.

If pm-go isn't on PATH, install it once (idempotent — re-running upgrades):

curl -fsSL https://raw.githubusercontent.com/alex-reysa/pm-go/main/scripts/install.sh | bash

Agentic Operator (default)

The operator runs inside the Claude Agent SDK and exposes a typed MCP tool surface so the model can drive pm-go without raw curl. Every tool call lands on agent_tool_calls (with FK to agent_runs.plan_id) for full replay.

Tool surface (13 tools):

• Lifecycle — pmgo_ensure_stack, pmgo_recover, pmgo_stop, pmgo_doctor, pmgo_status
• Submission — pmgo_submit_spec (rejects model-supplied specPath / repoRoot outside operator scope)
• Planning — pmgo_decompose_spec (despite the name, this starts the normal full-spec planning workflow via POST /plans)
• Manifest hooks — pmgo_update_manifest, pmgo_plan_first (handler-backed hooks; use the CLI Layer-A flow unless production handlers are present)
• Drive — pmgo_drive_plan (rejects approve override; only operator's --approve policy wins)
• Approve — pmgo_approve_pending (returns requires_confirmation when operator chose --approve none)
• Inspect — pmgo_why, pmgo_tail_events

Path containment, approval-policy enforcement, and FK-safe planId extraction are wired into the tool wrappers. pmgo_decompose_spec intentionally does NOT record planId on the tool-call row: POST /plans returns the id synchronously, but the plans row is async-inserted by persistPlan, so recording the FK immediately would race.

Layer-A: Decompose Large Specs

For specs > ~100 lines or scoping multiple milestones, use the Layer-A CLI to split the spec into an ordered MilestoneManifest before driving implementation. The command talks to a running API, so start the stack first:

# Step 1 — start the control plane and keep it running in one terminal
pm-go run --repo .

# Step 2 — from another shell, decompose, optionally edit, then plan the first milestone
pm-go decompose --repo . --spec ./feature.md --edit

# Step 3 — drive the plan id printed by decompose
pm-go drive --plan <plan-id>

pm-go decompose submits the spec, waits for the manifest, prints it, and unless --manifest-only is set, starts plan-first for the first milestone and prints the resulting plan id. Add --edit when the operator should review or adjust the manifest in $VISUAL / $EDITOR / vi before plan-first. Use --manifest-only when the next step is human review rather than immediate planning.

Provenance (decompositionId, milestoneId) flows through to the plans row so a downstream operator can trace the plan back to the manifest entry that scoped it. Current Layer-A behavior plans the first milestone only; it does not expose --plan-first or --milestone-id CLI flags, and it does not auto-chain later milestones.

Visibility

Four commands cover most diagnostic needs without raw tctl or DB queries:

pm-go doctor             # env + CLIs + auth + infra (postgres/temporal/migrations/ports)
pm-go doctor --repair    # also auto-fix what it can (docker up, db:migrate, mkdir)
pm-go status             # worker config, API /health, open Temporal workflows
pm-go why <id>           # one-sentence state + next action for any plan/phase/task/spec id

Use pm-go status first when something looks stuck — it tells you the configured task queue + namespace + open workflows. A "scheduled but never picked up" workflow is almost always one of: stale dist/ from before a code change, mismatched TEMPORAL_TASK_QUEUE, or the worker pointing at the wrong Temporal cluster.

Picking a Model

pm-go defaults to claude-opus-4-7 for every role. Override per-role or globally with env vars:

PM_GO_MODEL=claude-sonnet-4-6 pm-go --repo . --spec ./feature.md   # all worker roles
PLANNER_MODEL=claude-opus-4-7 IMPLEMENTER_MODEL=claude-sonnet-4-6 pm-go --repo . --spec ./feature.md

Recognized vars: PLANNER_MODEL, IMPLEMENTER_MODEL, REVIEWER_MODEL, PHASE_AUDITOR_MODEL, COMPLETION_AUDITOR_MODEL. PM_GO_MODEL is the shared fallback.

You should not need to edit packages/planner/src/*.ts — the env vars are the supported override.

When to Use Sub-Commands Instead of the Default Agent

The default pm-go (agent) covers the common case. Reach for sub-commands when:

pm-go run is the supervisor without the auto-drive. pm-go drive assumes the API is already up on :3001 and a plan exists in Postgres.

Spec Format

The spec is a Markdown file. Key sections:

• Objective — one-paragraph scope
• Scope — what's in / out
• Constraints — technical bounds (no new dependencies, must preserve API X, etc.)
• Acceptance Criteria — bullet list, each becomes a test target the reviewer enforces
• Repo Hints — files and directories the planner should focus on

See examples/spec-input-template.md and examples/golden-path/spec.md in the pm-go repo for canonical examples.

One spec → one plan → one repo. If the work spans multiple repos (e.g. a candidate API in repo A and a router in repo B), file two specs. Specs accept an explicit repoRoot per submission, so a single API can host plans for multiple repos.

State Machine

spec → plan → [for each phase] execute tasks → review → fix? → integrate → audit → complete → release

• Tasks within a phase run in parallel where dependencies allow.
• Review produces a ReviewReport with pass | changes_requested | blocked. changes_requested triggers up to maxReviewFixCycles repair cycles before escalation.
• Phase audit runs after task integration; completion audit runs after all phases complete.
• Release is gated on a passing completion audit.

pm-go drive runs all of this for you. The TUI (pnpm tui from the pm-go repo) is useful for human inspection but is not required.

Approval Gates

High-risk tasks/phases pause for explicit approval. In the default agent, --approve interactive is the default and prompts the operator; --yes defaults the agent to --approve all for autonomous runs. In pm-go drive and legacy pm-go implement --legacy-drive, --approve all is the default. --approve none exits when an approval is needed; resume by approving and re-running drive.

When legacy pm-go implement --legacy-drive pauses for approval, it leaves the API + worker up and prints the exact approval URL. Resolve the approval, then re-run pm-go drive --plan <uuid>.

Recovery

If the supervisor died mid-flight, or you don't know what state a plan/phase/task is in, run:

pm-go why <id>          # id can be plan, phase, task, or spec-document UUID

It returns one sentence with the current state and the exact next action. This is the single best diagnostic — reach for it before grepping source, opening the TUI, or hitting Postgres directly. If why says the plan is mid-phase, follow up with pm-go drive --plan <uuid> to resume from where the supervisor left off.

Common Failure Modes

• Worker workflow stuck "scheduled but never picked up" — almost always stale dist/. Run pnpm -r build from the pm-go repo, then restart the worker. pm-go status confirms task queue + namespace alignment.
• Diff-scope violation — the implementer wrote files outside the task's fileScope. Either widen the scope (requires a plan re-draft) or tighten the task description.
• Content-filter rejection — agent_runs.error_reason="ContentFilterError" and the task is blocked. Adjust the spec wording for the affected task and re-run.
• Phase won't advance — every task in the phase must be ready_to_merge. Find the laggards: curl http://localhost:3001/plans/<id> and look for tasks in reviewing / fixing.
• Workflow-id collision on resume — drive reports WorkflowExecutionAlreadyStarted after a supervisor restart. Run pm-go why <id> and pm-go status first, sweep stale process state with pm-go recover if needed, then re-run pm-go drive --plan <id>.
• Legacy pm-go implement --legacy-drive exits before the plan exists in DB — the supervisor's plan-persistence poll defaults to 45 minutes and reports whether the Temporal workflow is still running when it times out. If hit, keep the API + worker up, run pm-go status, and once GET /plans/<id> returns the plan envelope (which it does as soon as the planning row is inserted; see "Planning-row visibility" below) resume with pm-go drive --plan <id>. Use --plan-wait 60m for larger specs.

Planning-row visibility

GET /plans/:id returns the plan envelope as soon as the planning row is inserted — it does NOT 404 during planning. Older runbooks treated a 404 as the planning-pending signal; that has not been the behaviour since v0.8.8.x. The correct signal that a plan is still being authored is plans.status='draft' and an empty (or partial) plans.phases array in the envelope, NOT an HTTP 404. Use pm-go why <plan-id> rather than re-polling for an HTTP 200 to determine planning progress.

Release-evidence vocabulary

Three classes of evidence travel with a release; treat each class differently:

• Pre-release evidence MAY block the completion audit. Includes every required phase-audit report, every review report for a required acceptance criterion, the plan-level diff, and pre-release browser smoke when it can run against an integration worktree or pre-release preview. Missing pre-release evidence is the auditor's job to flag.
• Release artifacts never block. The PR summary (pr_summary) and the completion evidence bundle (completion_evidence_bundle) are produced BY FinalReleaseWorkflow AS the release lands. Their absence at audit time is expected, not a gap.
• Post-release evidence never blocks completion audit. Post-release smoke transcripts, deploy confirmations, and PR/merge metadata are recorded after release via the post_release_evidence artifact path (see the new recordPostReleaseEvidence activity). They feed traceability and operator dashboards, not the audit verdict. An operator who wants to gate release on post-release signal does so via completion_audit_reports.override_reason / overridden_by / overridden_at, NOT by waiting for post-release evidence inside the audit.

Do not advise operators to delay completion audit until post-release smoke arrives, and do not interpret a missing post_release_evidence artifact as a completion-audit blocker.

[pm-go] port <port> is held by another service

Phase-1 commands (status, drive, why, recover, run, implement, decompose) probe the API's identity endpoint before driving against it. When the probe gets HTTP back but the response isn't a pm-go identity envelope ({"service":"pm-go-api",...}), the CLI bails with a PmGoIdentityMismatchError whose first line is the stable, greppable prefix:

[pm-go] port 3001 is held by another service

Cause. Something other than the pm-go API is bound to the configured port — most often another tsx-loaded server squatting on 3001 (a stale apps/api process from a prior dev session, an unrelated local service, or a second pm-go instance that claimed the port first).

Remediation (two options).

1. Re-run the command with --port <other> to point pm-go at a free port.
2. Stop the conflicting process (pm-go ps then pm-go stop if it's a tracked pm-go supervisor; otherwise lsof -i :3001 and shut down the offending service yourself).

API Surface (advanced)

pm-go drive orchestrates the API for you. If the user is explicitly debugging API behavior or writing tests, read references/api-surface.md for the current raw endpoints. Do not copy old snippets from memory: POST /spec-documents uses body, and POST /plans needs both specDocumentId and repoSnapshotId.

Stub Mode (CI / fast smoke)

Set --runtime stub for fixture-driven runs that exercise the full pipeline without calling Claude:

pm-go --repo . --runtime stub --spec ./examples/golden-path/spec.md

Useful for smoke-testing pm-go itself, not for solving real specs. Stub fixtures live under packages/sample-repos/ and examples/.

Anti-Patterns

• Don't pre-check ANTHROPIC_API_KEY. OAuth alone is sufficient. Ask the supervisor instead — it prints which source it picked at boot.
• Don't edit packages/planner/src/*.ts to swap models. Use PM_GO_MODEL or per-role env vars.
• Don't invent Layer-A milestone-selection flags. The public CLI plans the first milestone only; verify pm-go decompose --help before documenting new flags.
• Don't pkill -f pm-go — it kills the operator's monitors too. Use pm-go ps to inspect the supervisor-owned process registry, then pm-go stop to shut them down cleanly. (Both commands target only pm-go's tracked PIDs, so editor and dev-server processes survive.)
• Don't run two supervisors against the same Postgres + Temporal. Port 3001 collides. pm-go status will tell you something is already on that port.
• Don't push to main of the pm-go repo while running pm-go on its own repo. Baseline drift causes false phase-audit scope violations on every phase. See docs/dogfood/v0.8.6-run.md for the cautionary tale.

Reference

• examples/golden-path/ — full walkthrough of a stub-mode run
• docs/getting-started.md — manual API flow
• docs/runbooks/ — operational recovery playbooks
• references/api-surface.md — raw API fallback for advanced/manual flows
• packages/contracts/src/ — domain types (Plan, Task, AgentRun, ReviewReport)
• artifacts/plans/ — generated plan Markdown after SpecToPlanWorkflow finishes

Source: alex-reysa/pm-go — distributed by TomeVault.