@unbrained/pm-cli 2026.9.25 → 2026.9.27

package/plugins/pm-claude/skills/pm-audit/SKILL.md

@@ -1,110 +1,21 @@
 ---
 name: pm-audit
-description: Audit pm-cli repositories with native pm MCP tools — health checks, validation, duplicate detection, aggregation, privacy review, and workflow health. Use when performing broad repository audits, release readiness checks, or agent-workflow health reviews.
+description: Use when asked to inspect pm tracker health, duplicate work, graph quality, privacy, or release readiness. Report verified findings and reuse existing lineage.
+license: MIT
 ---
 
-# pm Audit
+# pm Review
 
-Use for comprehensive repository audits, consistency reviews, and health diagnostics.
+Read the requested scope and the full live records behind any status claim.
+Use pm_health, pm_validate, pm_search, pm_graph, and pm_contracts as needed.
+Use pm_run for duplicate checks when no narrow tool covers them. Check omission
+receipts and follow cursors before reporting totals.
 
-## Audit Flow
+Classify each finding by its observable impact. Search all statuses before
+filing it and link it to its source with a meaningful typed relationship.
+Record exact commands and results on the canonical item. Keep credentials,
+private operational data, and raw meeting content out of public records.
 
-1. **Context snapshot** — `pm_context` with `depth: "standard"`.
-2. **Search for existing audits** — `pm_search` to avoid duplicate tracking.
-3. **Create or claim audit item** — then call `TaskCreate` to show in Claude Code's task panel.
-4. **Health check** — `pm_health` for tracker diagnostics.
-5. **Validation suite** — `pm_validate` with all checks enabled.
-6. **Aggregate** — `pm_run` with `action: "aggregate"` for governance view.
-8. **Calendar** — `pm_run` with `action: "calendar"` for deadline view.
-9. **Convert findings** — create pm items for each actionable finding.
-10. **Evidence** — `pm_comments` with verification summaries.
-11. **Close audit item** then `TaskUpdate(completed)`.
-
-## Hybrid TUI Sync
-
-After creating/claiming the audit tracking item:
-```
-TaskCreate:
-  subject: "[pm-xxxx] Audit: pm tracker health"
-  description: "Tracking full audit run — pm-xxxx"
-  activeForm: "Running pm audit"
-```
-Save the `taskId`. Call `TaskUpdate(in_progress)`.
-When closing: `TaskUpdate(completed)`.
-
-For each **finding item** created during the audit, also create a matching `TaskCreate` if you plan to work on it this session.
-
-## MCP Calls
-
-### Full Audit Suite
-```json
-{ "tool": "pm_context", "args": { "options": { "depth": "standard", "limit": "20" } } }
-{ "tool": "pm_health", "args": {} }
-{ "tool": "pm_validate", "args": { "options": { "checkResolution": true, "checkHistoryDrift": true, "checkFiles": true, "scanMode": "tracked-all" } } }
-{ "tool": "pm_run", "args": { "action": "aggregate", "options": { "groupBy": "status,type" } } }
-{ "tool": "pm_run", "args": { "action": "calendar", "options": { "view": "week", "format": "markdown", "include": "deadlines,reminders" } } }
-{ "tool": "pm_run", "args": { "action": "stats" } }
-```
-
-### Activity Review
-```json
-{ "tool": "pm_run", "args": { "action": "activity", "options": { "limit": "50" } } }
-```
-
-### Contracts and Schema Check
-```json
-{ "tool": "pm_contracts", "args": {} }
-```
-
-## Finding Classification
-
-| Severity | Action |
-|----------|--------|
-| Blocker | Create pm item with priority 0, status open |
-| Warning | Create pm item with priority 1 |
-| Info | Append to existing audit item comments |
-| No issue | Log as evidence in audit comments |
-
-## Evidence Format
-
-```json
-{
-  "tool": "pm_comments",
-  "args": {
-    "id": "pm-xxxx",
-    "options": {
-      "add": "Audit evidence: health=ok, validate=ok (2 warnings), dedupe=5 candidates, items=681, tests=1087 passing."
-    }
-  }
-}
-```
-
-## Privacy
-
-Keep sensitive operational data (telemetry credentials, user PII, internal endpoints) out of public docs and tracked comments. Use `pm notes` for sensitive context instead of `pm comments`.
-
-## Progressive Disclosure and Token Discipline
-
-Load the smallest thing that answers the question. Costs are measured.
-
-| Need | Call | Cost |
-|------|------|------|
-| Pick or resume work | `pm_next`, or `pm_context` with `limit: 10` | ~2.1-2.5k tok |
-| Exact flags for one command | `pm_contracts` with `command: "<name>", flagsOnly: true` | ~1-3.4k tok |
-| The whole command surface with per-command ceilings | `pm_contracts` with `summary: true` | ~2.6k tok |
-| An unfamiliar capability family | `pm_run` with `action: "guide", topic: "<topic>", depth: "brief"` | ~0.6-1k tok |
-
-Guide topics are declared at runtime: `pm guide` prints the current index and
-`pm guide <topic> --depth brief` expands one topic. Do not copy the list.
-
-Never load `docs/COMMANDS.md` (~29k tok) or `docs/SDK.md` (~45k tok) whole.
-
-Bound every read in this order: projection (`outputInclude`), then row limit
-(`outputLimit` or `limit`), then `outputBudget`, then `outputCursor` to resume.
-Read the `omission_receipt` before treating a result as complete — a
-budget-truncated read is a claim about the part it withheld, and a truncated
-list must never be summarized as if it were the whole population.
-
-Author identity is detected automatically. Never pass `author` and never set
-`PM_AUTHOR`; the harness, model, effort, role, and topic are probed and recorded
-on every history entry.
+An inspection does not imply permission to rewrite history or fabricate
+relationships. When asked to fix findings, claim only the work being changed
+and follow the pm-developer skill.

package/plugins/pm-claude/skills/pm-audit/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "pm Review"
+  short_description: "Inspect tracker health and readiness"
+  default_prompt: "Review live pm health, quality, and release evidence; report verified findings."

package/plugins/pm-claude/skills/pm-developer/SKILL.md

@@ -1,149 +1,26 @@
 ---
 name: pm-developer
-description: Run the pm-cli developer execution loop — orient, claim, implement, link evidence, verify, and close — with native MCP tools. Use when coding, debugging, refactoring, or shipping changes tracked in pm items.
+description: Use when implementing, debugging, refactoring, or shipping code, tests, or docs tracked by pm. Keep SDK behavior and verification linked to the active item.
+license: MIT
 ---
 
 # pm Developer
 
-Use this skill for implementation work that changes code, docs, tests, or release artifacts.
-
-## Canonical Execution Loop
-
-1. **Orient** — `pm_context` then `pm_search` for relevant existing items.
-2. **Decide** — reuse an existing item when one matches; create only when needed.
-3. **Claim** — `pm_claim` before any substantial edits.
-4. **Sync TUI** — call `TaskCreate` to mirror the pm item in Claude Code's task panel (see Hybrid TUI Sync).
-5. **Implement** — make the changes, link files/docs/tests as you go.
-6. **Verify** — run linked tests plus local quality gates.
-7. **Evidence** — `pm_comments` with what changed and what passed.
-8. **Close** — `pm_close` then `pm_release` then `TaskUpdate(completed)`.
-
-## Hybrid TUI Sync
-
-pm is the **persistent store**. Claude Code's task panel is the **live session view**.
-
-### After pm_claim (or pm_create when starting fresh)
-
-```
-TaskCreate:
-  subject: "[pm-xxxx] <item title>"
-  description: "Tracking pm item pm-xxxx. AC: <acceptance_criteria>"
-  activeForm: "Implementing pm-xxxx"
-```
-
-Save the returned `taskId`. Then:
-
-```
-TaskUpdate: { taskId: <saved>, status: "in_progress" }
-```
-
-### After pm_close + pm_release
-
-```
-TaskUpdate: { taskId: <saved>, status: "completed" }
-```
-
-## Step-by-Step MCP Calls
-
-### 1. Orient
-
-```json
-{ "tool": "pm_context", "args": { "options": { "limit": "10" } } }
-{ "tool": "pm_search", "args": { "query": "task keywords", "options": { "limit": "10" } } }
-```
-
-### 2. Claim
-
-```json
-{ "tool": "pm_claim", "args": { "id": "pm-xxxx" } }
-{ "tool": "pm_update", "args": { "id": "pm-xxxx", "options": { "status": "in_progress" } } }
-```
-
-Then immediately: `TaskCreate` → `TaskUpdate(in_progress)` as shown above.
-
-### 3. Link Evidence (during implementation)
-
-```json
-{ "tool": "pm_files", "args": { "id": "pm-xxxx", "options": { "add": ["path=src/file.ts,scope=project,note=implementation"] } } }
-{ "tool": "pm_docs", "args": { "id": "pm-xxxx", "options": { "add": ["path=docs/GUIDE.md,scope=project,note=updated"] } } }
-{ "tool": "pm_test", "args": { "id": "pm-xxxx", "options": { "add": ["command=node scripts/run-tests.mjs test -- tests/unit/foo.spec.ts,scope=project,timeout_seconds=240"] } } }
-```
-
-### 4. Validate
-
-```json
-{ "tool": "pm_validate", "args": { "options": { "checkResolution": true, "checkHistoryDrift": true } } }
-```
-
-### 5. Add Evidence Comment
-
-```json
-{ "tool": "pm_comments", "args": { "id": "pm-xxxx", "options": { "add": "Evidence: changed src/foo.ts. All tests pass. pm validate ok." } } }
-```
-
-### 6. Close
-
-```json
-{ "tool": "pm_close", "args": { "id": "pm-xxxx", "reason": "Implementation complete. Acceptance criteria met." } }
-{ "tool": "pm_release", "args": { "id": "pm-xxxx" } }
-```
-
-Then: `TaskUpdate: { taskId: <saved>, status: "completed" }`
-
-## Verification Defaults
-
-Run these before closing any implementation item:
-
-- `pnpm build` (or project build command)
-- `node scripts/run-tests.mjs test -- <target>` (targeted test)
-- `node scripts/run-tests.mjs coverage` (coverage gate)
-- `pm_validate` with `checkResolution: true`
-
-## Create When Needed
-
-When no existing item matches:
-
-```json
-{
-  "tool": "pm_create",
-  "args": {
-    "options": {
-      "title": "Fix: concise description of the problem",
-      "description": "Root cause and approach.",
-      "type": "Task",
-      "status": "open",
-      "priority": "1",
-      "tags": "relevant,tags",
-      "createMode": "progressive"
-    }
-  }
-}
-```
-
-Then call `TaskCreate` + `TaskUpdate(in_progress)` after claiming.
-
-## Progressive Disclosure and Token Discipline
-
-Load the smallest thing that answers the question. Costs are measured.
-
-| Need | Call | Cost |
-|------|------|------|
-| Pick or resume work | `pm_next`, or `pm_context` with `limit: 10` | ~2.1-2.5k tok |
-| Exact flags for one command | `pm_contracts` with `command: "<name>", flagsOnly: true` | ~1-3.4k tok |
-| The whole command surface with per-command ceilings | `pm_contracts` with `summary: true` | ~2.6k tok |
-| An unfamiliar capability family | `pm_run` with `action: "guide", topic: "<topic>", depth: "brief"` | ~0.6-1k tok |
-
-Guide topics are declared at runtime: `pm guide` prints the current index and
-`pm guide <topic> --depth brief` expands one topic. Do not copy the list.
-
-Never load `docs/COMMANDS.md` (~29k tok) or `docs/SDK.md` (~45k tok) whole.
-
-Bound every read in this order: projection (`outputInclude`), then row limit
-(`outputLimit` or `limit`), then `outputBudget`, then `outputCursor` to resume.
-Read the `omission_receipt` before treating a result as complete — a
-budget-truncated read is a claim about the part it withheld, and a truncated
-list must never be summarized as if it were the whole population.
-
-Author identity is detected automatically. Never pass `author` and never set
-`PM_AUTHOR`; the harness, model, effort, role, and topic are probed and recorded
-on every history entry.
+Orient with pm-workflow before editing. Claim the canonical item and keep its
+scope aligned with the change.
+
+1. Read the complete source owner and relevant tests before editing. For a
+   regression, identify the observable contract and the change that would
+   make its test fail.
+2. Put domain behavior in the public pm SDK. Keep CLI and MCP adapters thin,
+   with transport-specific tests only where those boundaries add risk.
+3. Link changed files, docs, and isolated runnable tests through pm_files,
+   pm_docs, and pm_test.
+4. Run the focused test, build, full coverage, static quality, and installed
+   consumer checks required by the affected contract. Keep tracker test roots
+   separate from the real repository.
+5. Record exact results in pm_comments. Close with structured resolution,
+   expected, and actual fields only when the criteria are actually met.
+
+Prefer pm_contracts for flags and the SDK public surface for integration APIs.
+An unknown review provider or missing check is not approval.

package/plugins/pm-claude/skills/pm-developer/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "pm Developer"
+  short_description: "Deliver code with linked pm evidence"
+  default_prompt: "Use pm to claim, implement, verify, document, and close this repository change."

package/plugins/pm-claude/skills/pm-planner/SKILL.md

@@ -1,173 +1,22 @@
 ---
 name: pm-planner
-description: Plan and organize pm CLI work — triage requests, break epics into tasks, prioritize the backlog, and keep the tracker clean. Use when decomposing features, planning sprints, or organizing incoming requests.
+description: Use when triaging requests, organizing a backlog, decomposing an epic, or making a durable pm plan. Search all statuses and build evidence-backed typed relationships.
+license: MIT
 ---
 
 # pm Planner
 
-Use this skill when decomposing features, organizing backlogs, triaging incoming work, or planning multi-item projects.
-
-## Planning Loop
-
-1. **Survey** — `pm_context` then `pm_list` with no filter to see the full active backlog.
-2. **Deduplicate** — `pm_search` before every create to avoid duplicate items.
-3. **Decompose** — break large items into Epic → Feature → Task hierarchy.
-4. **Prioritize** — use `pm_update` to set priority (`0`=critical … `4`=minimal).
-5. **Link parents** — always set `parent` on child items via `pm_update`.
-6. **Sync TUI** — call `TaskCreate` for any item you claim during this session (see Hybrid TUI Sync).
-7. **Validate** — `pm_validate` after batch creates to check consistency.
-
-## Hybrid TUI Sync
-
-pm is the **persistent store**. Claude Code's task panel is the **live session view**.
-
-When you claim a planning item or start active decomposition work:
-```
-TaskCreate:
-  subject: "[pm-xxxx] Plan: <epic/feature title>"
-  description: "Planning pm-xxxx — decomposing into child items"
-  activeForm: "Planning pm-xxxx"
-```
-Save the `taskId`. Call `TaskUpdate(in_progress)` when active.
-Call `TaskUpdate(completed)` when the planning item is closed.
-
-## MCP Call Patterns
-
-### Survey the Backlog
-```json
-{ "tool": "pm_context", "args": { "options": { "depth": "standard", "limit": "20" } } }
-{ "tool": "pm_list", "args": { "options": { "limit": "50" } } }
-```
-
-### Deduplicate Before Creating
-```json
-{ "tool": "pm_search", "args": { "query": "feature keywords", "options": { "limit": "10" } } }
-```
-
-### Create Epic → Feature → Task Hierarchy
-```json
-{
-  "tool": "pm_create",
-  "args": {
-    "options": {
-      "title": "Epic: Authentication system overhaul",
-      "type": "Epic",
-      "status": "open",
-      "priority": "1",
-      "description": "High-level initiative description.",
-      "createMode": "progressive"
-    }
-  }
-}
-```
-Then for each child feature/task, set `parent: "pm-xxxx"` via `pm_update`.
-
-### Prioritize the Backlog
-
-```json
-{ "tool": "pm_list", "args": { "options": { "status": "open", "limit": "30" } } }
-{ "tool": "pm_update", "args": { "id": "pm-xxxx", "options": { "priority": "0" } } }
-```
-
-### Link Parent
-
-```json
-{ "tool": "pm_update", "args": { "id": "pm-child", "options": { "parent": "pm-epic" } } }
-```
-
-### Run Dedupe Audit After Batch Work
-
-```json
-```
-
-### Validate After Batch Creates
-
-```json
-{ "tool": "pm_validate", "args": { "options": { "checkResolution": true } } }
-```
-
-## Item Type Guide
-
-| Type | When to use |
-|------|------------|
-| `Epic` | Large initiative spanning multiple features (weeks/months) |
-| `Feature` | Distinct capability under an epic (days) |
-| `Task` | Atomic unit of implementation work (hours) |
-| `Bug` | Defect with reproduce steps |
-| `Story` | User-facing narrative unit |
-
-## Priority Reference
-
-- `0` = critical — blocking release or other work
-- `1` = high — important, should be done soon
-- `2` = normal — standard priority (default)
-- `3` = low — nice to have
-- `4` = minimal — defer unless time allows
-
-## Acceptance Criteria Best Practice
-
-Write AC as numbered, testable assertions:
-1. Given [precondition], when [action], then [result].
-2. All existing tests pass.
-3. Coverage gate maintained.
-
-Set via `pm_create` with `acceptanceCriteria` field or `pm_update` with `options.acceptanceCriteria`.
-
-## Built-in Plan workflow (`pm_plan` / `pm plan`)
-
-For durable Codex-style ExecPlans, Claude-style Plan Mode, and Cursor-style editable checklists, pm exposes a first-class `Plan` item type plus a `pm plan` command family. Use [Command Reference: Plan Workflow](../../../../docs/COMMANDS.md#plan-workflow) as the canonical lifecycle recipe; this skill keeps only Claude-specific routing notes.
-
-### When to use `pm plan` vs. `pm_create` + tasks
-
-- `pm plan` — when you need a **living, resumable plan** with ordered steps, evidence, decisions, discoveries, validation, and the option to later materialize selected steps as real Tasks. Best for plan-then-execute workflows.
-- `pm_create` (Epic/Feature/Task) — when the work is already decomposed and you just need persistent backlog items.
-
-### Claude MCP routing
-
-Call the same lifecycle through the `pm_plan` MCP tool. Use `harness: "claude-code"` on creation, `subcommand: "show"` with `depth: "brief"` for low-token reads, and `subcommand: "resume"` after compaction or handoff.
-
-### Harness mapping cheatsheet
-
-| Harness signal | pm_plan equivalent |
-|----------------|--------------------|
-| Codex `update_plan` step status | `subcommand=update-step` with `stepStatus` |
-| Claude `TaskCreate` | `subcommand=add-step` |
-| Claude `TaskUpdate(in_progress)` | `subcommand=update-step` with `stepStatus=in_progress` |
-| Cursor edit a plan step | `subcommand=update-step` |
-| Resume after compaction | `subcommand=resume` with `resumeContext` then `subcommand=show --depth deep` |
-| Approve before edits | `subcommand=approve` |
-| Convert checklist into real items | `subcommand=materialize` |
-
-## Safety
-
-- Never pass `path` during real repository tracking.
-- Always `pm_search` before `pm_create`/`pm plan create` — avoid duplicates.
-- Run `pm_validate` after batch changes.
-- Set exactly one step `in_progress` per plan; pass `allowMultipleActive: true` only for explicit parallel branches.
-- Use `pm_plan` `subcommand=resume` after long-running sessions so the next agent can pick up with a deterministic context.
-
-## Progressive Disclosure and Token Discipline
-
-Load the smallest thing that answers the question. Costs are measured.
-
-| Need | Call | Cost |
-|------|------|------|
-| Pick or resume work | `pm_next`, or `pm_context` with `limit: 10` | ~2.1-2.5k tok |
-| Exact flags for one command | `pm_contracts` with `command: "<name>", flagsOnly: true` | ~1-3.4k tok |
-| The whole command surface with per-command ceilings | `pm_contracts` with `summary: true` | ~2.6k tok |
-| An unfamiliar capability family | `pm_run` with `action: "guide", topic: "<topic>", depth: "brief"` | ~0.6-1k tok |
-
-Guide topics are declared at runtime: `pm guide` prints the current index and
-`pm guide <topic> --depth brief` expands one topic. Do not copy the list.
-
-Never load `docs/COMMANDS.md` (~29k tok) or `docs/SDK.md` (~45k tok) whole.
-
-Bound every read in this order: projection (`outputInclude`), then row limit
-(`outputLimit` or `limit`), then `outputBudget`, then `outputCursor` to resume.
-Read the `omission_receipt` before treating a result as complete — a
-budget-truncated read is a claim about the part it withheld, and a truncated
-list must never be summarized as if it were the whole population.
-
-Author identity is detected automatically. Never pass `author` and never set
-`PM_AUTHOR`; the harness, model, effort, role, and topic are probed and recorded
-on every history entry.
+1. Read pm_context and the relevant parent or outcome milestone.
+2. Search all statuses and inspect nearest matches before creating anything.
+3. Use a Plan when ordered steps, decisions, discoveries, and handoff must
+   remain resumable. Use Epic, Feature, Task, Issue, and Decision items for
+   durable backlog work with different ownership.
+4. Give new work a parent, testable acceptance criteria, expected result, and
+   a create-time duplicate-check explanation.
+5. Add typed edges only when evidence supports their meaning. Use dependency
+   edges for ordering and verification edges for proof; do not invent graph
+   depth or fill the graph with related edges.
+6. Validate the resulting tracker and summarize the next actionable items.
+
+Use pm_plan for a living plan and pm_contracts to discover exact options.
+Keep one plan step in progress unless parallel work is explicitly intended.

package/plugins/pm-claude/skills/pm-planner/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "pm Planner"
+  short_description: "Plan and connect project work"
+  default_prompt: "Use pm to deduplicate this request, plan its work, and link it to the project outcome."

package/plugins/pm-claude/skills/pm-release/SKILL.md

@@ -1,108 +1,24 @@
 ---
 name: pm-release
-description: Run pm-cli release-readiness workflows — validation, coverage gates, changelog, GitHub checks — using native MCP tools with linked evidence. Use when preparing releases, publishing, or verifying post-release state.
+description: Use when preparing a pm release, reviewing its gates, publishing, or verifying the released artifact. Keep source, reviewed head, main, tag, registry, and telemetry evidence distinct.
+license: MIT
 ---
 
 # pm Release
 
-Use for release preparation, publication, and post-release verification.
-
-## Release Loop
-
-1. **Find or create** the release pm item (after duplicate check with `pm_search`).
-2. **Claim** it — then call `TaskCreate` to show the release in Claude Code's task panel.
-3. **Run local gates** before tagging:
-   - `pnpm build` — clean build
-   - `node scripts/run-tests.mjs coverage` — full coverage gate
-   - `pm_validate` with all checks
-   - `pm_health` for tracker status
-4. **Tag and push** — let CI run.
-5. **Verify GitHub checks** — `gh run list --limit 5` after push.
-6. **Record evidence** with `pm_comments`.
-7. **Close and release** the pm item, then `TaskUpdate(completed)`.
-
-## Hybrid TUI Sync
-
-### After claiming the release item
-```
-TaskCreate:
-  subject: "[pm-xxxx] Release: <version>"
-  description: "Release gate tracking for pm-xxxx"
-  activeForm: "Running release gates"
-```
-Save the `taskId`. Then `TaskUpdate(in_progress)`.
-
-### After pm_close + pm_release
-```
-TaskUpdate: { taskId: <saved>, status: "completed" }
-```
-
-## MCP Calls
-
-### Pre-Release Validation
-```json
-{ "tool": "pm_validate", "args": { "options": { "checkResolution": true, "checkHistoryDrift": true, "checkFiles": true } } }
-{ "tool": "pm_health", "args": {} }
-{ "tool": "pm_context", "args": { "options": { "depth": "standard" } } }
-```
-
-### Link Release Artifacts
-```json
-{ "tool": "pm_docs", "args": { "id": "pm-xxxx", "options": { "add": ["path=CHANGELOG.md,scope=project,note=release-notes"] } } }
-{ "tool": "pm_files", "args": { "id": "pm-xxxx", "options": { "add": ["path=package.json,scope=project,note=version"] } } }
-```
-
-### Record Evidence
-```json
-{
-  "tool": "pm_comments",
-  "args": {
-    "id": "pm-xxxx",
-    "options": {
-      "add": "Release evidence: build ok. Tests pass. 100% coverage. pm validate ok. GitHub CI green. npm publish ok."
-    }
-  }
-}
-```
-
-## Gate Script Reference
-
-```bash
-pnpm build                                   # TypeScript build
-node scripts/run-tests.mjs coverage          # Full test + coverage
-node scripts/release/run-gates.mjs           # All release gates
-gh run list --limit 5                        # GitHub CI status
-gh run view <run-id>                         # Detailed run status
-```
-
-## Safety
-
-- Never set `path` for real repository tracking — only for sandbox tests.
-- Run coverage gate before any `npm publish` or version tag.
-- Confirm GitHub CI green before closing the release item.
-
-## Progressive Disclosure and Token Discipline
-
-Load the smallest thing that answers the question. Costs are measured.
-
-| Need | Call | Cost |
-|------|------|------|
-| Pick or resume work | `pm_next`, or `pm_context` with `limit: 10` | ~2.1-2.5k tok |
-| Exact flags for one command | `pm_contracts` with `command: "<name>", flagsOnly: true` | ~1-3.4k tok |
-| The whole command surface with per-command ceilings | `pm_contracts` with `summary: true` | ~2.6k tok |
-| An unfamiliar capability family | `pm_run` with `action: "guide", topic: "<topic>", depth: "brief"` | ~0.6-1k tok |
-
-Guide topics are declared at runtime: `pm guide` prints the current index and
-`pm guide <topic> --depth brief` expands one topic. Do not copy the list.
-
-Never load `docs/COMMANDS.md` (~29k tok) or `docs/SDK.md` (~45k tok) whole.
-
-Bound every read in this order: projection (`outputInclude`), then row limit
-(`outputLimit` or `limit`), then `outputBudget`, then `outputCursor` to resume.
-Read the `omission_receipt` before treating a result as complete — a
-budget-truncated read is a claim about the part it withheld, and a truncated
-list must never be summarized as if it were the whole population.
-
-Author identity is detected automatically. Never pass `author` and never set
-`PM_AUTHOR`; the harness, model, effort, role, and topic are probed and recorded
-on every history entry.
+Inspect the release item and current release position before mutation. Search
+for an existing release or incident lineage and reuse it.
+
+1. Link the changed files, changelog, compatibility evidence, and release
+   tests to the item that owns them.
+2. Run the build, full coverage, static quality, security, docs, package,
+   version, and hosted analysis gates required by the repository.
+3. Verify the exact reviewed PR head and resolve actionable feedback before
+   merging. A silent or quota-limited provider is unavailable evidence.
+4. After publishing, verify the main commit, tag, GitHub Release, npm
+   package, installed npm and Bun consumers, Sentry, and telemetry separately.
+5. Record immutable close evidence and release the claim.
+
+Use the repository release pipeline and pm_contracts rather than a remembered
+flag spelling. Never assume a successful API request proves a scheduled
+release or physical telemetry delivery.

package/plugins/pm-claude/skills/pm-release/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "pm Release"
+  short_description: "Verify release gates and artifacts"
+  default_prompt: "Use pm to prepare, gate, and verify this release with distinct source and published evidence."