Runtime Components

Table of Contents

• What Runtime Components Are
• Sub-Agents Catalog
• Hooks Catalog
• Output Styles Catalog
• How to Discover and Invoke
• Cross-Client Compatibility
• Composition Patterns

What Runtime Components Are

pm-skills ships two layers of capability:

Content library (the bulk of the repo): 68 skills, 11 slash commands, 12 workflows, 27 enforcing CI validators. Skills are content - reference material an AI reads at invocation time. They do not execute logic on their own.

Runtime components (this catalog): Plugin features that take action. Sub-agents are dispatched by Claude Code’s intent classifier and run in their own context window. Hooks fire on lifecycle events (PreToolUse, PostToolUse, Stop). Output styles transform how Claude formats responses.

The distinction matters because skills are portable across clients (any client that supports the agentskills.io specification can read SKILL.md files), but runtime components are Claude Code plugin-specific. pm-skills delivers sub-agent intent to non-Claude clients via dispatch skills (see Cross-Client Compatibility) so users on Codex CLI, Cursor, Windsurf, Copilot, or Gemini CLI lose no functional access.

Layer	What it is	How AI uses it	Cross-client portable?
Skills (68)	Reference content with frontmatter, templates, examples	AI reads SKILL.md at invocation time	Yes (per agentskills.io spec)
Commands (11)	The /workflow-* orchestrators plus the /chain ad-hoc runner	User types /workflow-name or /chain, AI walks the skill sequence	Claude Code native
Workflows (12)	Multi-skill chains for full lifecycle moments	AI walks ordered skill invocations	Native; portable conceptually
Sub-agents (6; the v2.16 slate of 4 plus pm-workflow-orchestrator added v2.24.0 and pm-skill-router added v2.29.0)	Plugin components matched via description: and run in isolated context, or explicitly dispatched for maintainer tooling	Claude Code’s intent classifier delegates; OR user invokes the dispatch skill (/pm-skills:utility-pm-{role}) or @-mentions @agent-pm-skills:pm-{role}; pm-skill-router is explicit-only, maintainer-run	Claude Code only; dispatch skills provide cross-client parity (5 of 6; pm-skill-router has none)
Hooks (0; v2.17+ scope)	Lifecycle event handlers (PreToolUse, PostToolUse, Stop, etc.)	Claude Code fires automatically on configured events	Claude Code only
Output styles (0; v2.18+ scope)	Response formatting transforms	Claude applies when style is active	Claude Code only

Sub-Agents Catalog

v2.16.0 introduces sub-agents as the first runtime-component class. 4 sub-agents ship in this release. v2.24.0 adds a fifth, pm-workflow-orchestrator, and v2.29.0 adds a sixth, pm-skill-router (an internal tooling instrument, not a user-facing PM sub-agent), so the catalog below now lists 6 sub-agents. Each definition lives at agents/{name}.md, the fixed path Claude Code’s plugin runtime auto-discovers (the directory was named subagents/ through v2.16.x and renamed to agents/ in v2.17.0 W2; see project structure).

Sub-agent	Audience	Trigger	Lifetime	Tool Surface	Composition	Dispatch Skill
pm-critic	User (PM authoring artifacts)	Proactive after PM-artifact-producing skills (Claude Code) + explicit utility-pm-critic	Single turn	Read, Grep, Glob (no write; no Bash)	Skill-revise-recheck loop with deliver-prd, foundation-okr-writer, foundation-meeting-recap, foundation-persona, foundation-lean-canvas, discover-interview-synthesis, etc.	skills/utility-pm-critic/
pm-skill-auditor	User + Maintainer (audience straddles)	Explicit only: utility-pm-skill-auditor or @agent-pm-skills:pm-skill-auditor; chained from pm-release-conductor at gate G0	Multi-turn (may ask follow-up questions)	Bash, Read, Grep, Glob (no Edit; no Agent; detection-only)	Composes the enforcing validator suite via scripts/pre-tag-validate.{sh,ps1}; runs cross-cutting checks; re-derives aggregate counters	skills/utility-pm-skill-auditor/
pm-changelog-curator	Maintainer	Explicit only: utility-pm-changelog-curator or @agent-pm-skills:pm-changelog-curator; chained from pm-release-conductor at gate G2	Single turn (standalone or chained)	Bash, Read, Grep (no Edit; no Agent)	Composes with git log + CLAUDE.md hygiene rules; chained from conductor at G2	skills/utility-pm-changelog-curator/
pm-release-conductor	Maintainer	Explicit only: /pm-skills:utility-pm-release-conductor v{X.Y.Z} or @agent-pm-skills:pm-release-conductor (not proactive; safety from the 6 gates, not the invocation path)	Full release flow (often 30+ minutes)	Bash, Read, Edit, Grep, Glob, Agent (chain-permitted per _chain-permitted.yaml)	Chains to pm-skill-auditor at G0 + G2.5 and to pm-changelog-curator at G2; reads docs/contributing/release-runbook.md for gate definitions	skills/utility-pm-release-conductor/
pm-workflow-orchestrator	User + Maintainer	Explicit only: /pm-skills:utility-pm-workflow-orchestrator or @agent-pm-skills:pm-workflow-orchestrator (not proactive; safety from the per-step checkpoints, not the invocation path)	Multi-step run (pauses at each step for go/no-go)	Skill, Read, Grep, Glob, Bash, Edit (NO Agent; delegates to skills, not sub-agents; not chain-permitted)	Runs an ordered sequence of pm-skills against a saved foundation-prioritized-action-plan (Mode A) or a user-named chain (Mode B); inlines pm-critic’s leaf agent rather than nesting its dispatch skill	skills/utility-pm-workflow-orchestrator/
pm-skill-router	Maintainer (internal tooling instrument, not user-facing; AGENTS.md states this explicitly)	Explicit only: key-free --emit-tasks dispatch via the Agent tool (Haiku-pinned by default) from a maintainer session, or a direct Messages API call (--model=<id>, requires ANTHROPIC_API_KEY)	Single turn (one query in, one skill name or none out)	Read (no Write; no Bash)	Judges a query against the catalog by description: match only; the routing engine behind the enforcing scripts/check-new-skill-collision.mjs gate and the scripts/run-router-evals.mjs trigger-eval baseline, both maintainer-run repo-governance tools	None (no dispatch skill; see the Sub-Agent Compatibility Matrix)

For per-sub-agent cross-client status (Claude Code / Codex CLI / Cursor / Windsurf / Copilot CLI / Gemini CLI), see the canonical Sub-Agent Compatibility Matrix. Rows fill in as each sub-agent ships in its respective phase of subagents-integration-plan.md. Behavioral contracts live in the corresponding spec_pm-{name}.md files alongside the integration plan.

Sub-agent invariants

These invariants hold for every sub-agent in this catalog (enforced by scripts/validate-agents-md.{sh,ps1} extended per master plan D19):

• Definition file at agents/{name}.md with YAML frontmatter declaring name, description, tools, model, memory
• Dispatch skill at skills/utility-pm-{role}/ for cross-client access (the v2.16 companion-command pairing was retired in v2.22.0); pm-skill-router is the sole exception, an internal tooling instrument with no dispatch skill (see above)
• tools: frontmatter is a comma-separated scalar (per D20), e.g. tools: Read, Grep, Glob
• Agent in tools: requires entry in agents/_chain-permitted.yaml per D21 (HARD FAIL otherwise)
• Chain depth capped at 2 levels (D14)
• Severity grammar P0 / P1 / P2 / P3 across all findings output (D15)
• System prompt is referential, not duplicative (D12)
• Default memory: none unless escalation is justified (D13)

Hooks Catalog

No hooks ship in v2.16.0. The roadmap’s PostToolUse frontmatter hook (R-24) is deferred to v2.17 once sub-agents are battle-tested; the hook-triggered sub-agent invocation pattern (R-65) depends on the v2.16 sub-agent slate landing first.

Hook	Event	Status
(none in v2.16.0)	(none)	v2.17+ scope

Output Styles Catalog

No output styles ship in v2.16.0. The roadmap’s PM Voice output style (R-45) is deferred to v2.18+.

Style	Purpose	Status
(none in v2.16.0)	(none)	v2.18+ scope

How to Discover and Invoke

Native Claude Code path

Sub-agents are discovered automatically when the pm-skills plugin is installed. Two invocation paths exist:

Proactive (Claude Code matches description: and dispatches):

User: "Here's the PRD draft for review."
Claude Code: [intent classifier matches pm-critic's `use proactively` description]
             [dispatches to pm-critic in fresh context window]
             [pm-critic reads PRD, produces P0-P3 findings, returns to main context]
Claude: "I've reviewed your PRD. Three P1 findings to address..."

Only pm-critic ships with a proactive trigger in v2.16.0 per master plan D7. The 3 utility sub-agents (pm-skill-auditor, pm-changelog-curator, pm-release-conductor) are explicit-only.

Explicit (user invokes the dispatch skill):

User: utility-pm-critic docs/specs/my-prd.md
Claude Code: [invokes pm-critic with the artifact path as $ARGUMENTS]

Dispatch skills (one per sub-agent):

• utility-pm-critic [artifact path] invokes pm-critic
• utility-pm-skill-auditor invokes pm-skill-auditor
• utility-pm-changelog-curator [since-tag] invokes pm-changelog-curator
• utility-pm-release-conductor [version] invokes pm-release-conductor
• utility-pm-workflow-orchestrator [plan path] [--auto] [--force-auto] [--dry-run] invokes pm-workflow-orchestrator (added v2.24.0)

The orchestrator’s explicit @-mention form, mirroring the others:

User: @agent-pm-skills:pm-workflow-orchestrator _pm-skills/plans/storevine-q3.md
Claude Code: [invokes pm-workflow-orchestrator with the plan path as $ARGUMENTS]
             [parses Section 7 runnable blocks, runs them one at a time]
             [pauses for go/no-go before each step; stops on a failed or empty step]

Dispatch skill path (non-Claude clients)

On non-Claude clients (Codex CLI, Cursor, Windsurf, Copilot, Gemini CLI), sub-agents are not natively available. Users invoke the dispatch skill at skills/utility-pm-{role}/ which provides the same intent. The dispatch skill SKILL.md detects runtime and either:

1. (Claude Code path) invokes the native sub-agent via @agent-pm-skills:pm-{role} syntax, OR
2. (Non-Claude path) reads agents/pm-{role}.md and executes its system prompt inline

See Cross-Client Compatibility for details. Dispatch skill availability is conditional on Phase 2 spike outcomes; see the integration plan.

Cross-Client Compatibility

Per master plan D11 (amended) + D30:

Single-tool user assumption. pm-skills does NOT assume users run multiple AI tools simultaneously. Users typically pick ONE primary AI tool (Claude Code, Codex CLI, Cursor, Windsurf, Copilot, or Gemini CLI) for a given workflow. pm-skills delivers full value to all clients, but the SHAPE of the access differs:

Client	Skills access	Sub-agent access
Claude Code	Native plugin install; auto-discovery	Native plugin sub-agent at agents/{name}.md
Codex CLI	Sync to ~/.codex/skills/ per agentskills.io	Dispatch skill at skills/utility-pm-{role}/ reads agents/{name}.md inline
Cursor	Sync to .cursor/skills/	Same dispatch pattern
Windsurf	Sync to .windsurf/skills/	Same dispatch pattern
Copilot	Sync to .copilot/skills/	Same dispatch pattern
Gemini CLI	Sync to .gemini/skills/	Same dispatch pattern

Dispatch skill mechanism. Each dispatch skill is a thin pm-skills utility skill (~50 lines of SKILL.md) that contains conditional instructions:

If you are running in Claude Code with the pm-skills plugin: invoke @agent-pm-skills:pm-critic (or equivalent) on the target. If you are running in any other client: read the canonical agent definition at agents/pm-{name}.md and execute the system prompt body as your operating instructions for this turn.

The dispatch skill is portable per agentskills.io. The sub-agent definition file (agents/{name}.md) is portable as plain markdown - any AI that can read a file can execute its prompt body inline. There is no drift risk: the canonical system prompt lives in ONE place (the sub-agent definition), and the dispatch skill references it.

codex-rescue is NOT a baseline. Prior framing (pre-amendment) assumed users had Claude Code AND Codex CLI AND the codex-rescue plugin. None of these are universal. codex-rescue is an optional shortcut for the minority of users running both Claude Code and Codex CLI; it is not part of the dispatch path.

Functional gaps on non-Claude clients (acceptable per single-tool user assumption):

• Proactive triggering (pm-critic’s auto-fire after PM-artifact-producing skills) is a Claude Code frontmatter feature. On non-Claude clients, sub-agent invocation is explicit-only.
• User-level overrides via .claude/agents/ (Claude Code’s per-user agent dir) have no portable equivalent. Non-Claude users edit the dispatch skill SKILL.md locally instead.
• Plugin-level settings via .claude/pm-skills.local.md (deferred to v2.17+) are Claude-Code-specific.

These gaps are documented but not blocking; they reflect the platform differences, not pm-skills design choices.

Composition Patterns

pm-skills sub-agents compose with each other and with the skills library in specific patterns. The strategy doc at docs/internal/_working/subagents/subagent-strategy_2026-05-07.md catalogs all patterns; this section summarizes the v2.16.0 active ones.

Pattern 1: Skill -> sub-agent (proactive review loop)

A PM-artifact-producing skill (deliver-prd, foundation-okr-writer, foundation-meeting-recap, foundation-persona, foundation-lean-canvas, discover-interview-synthesis) emits its artifact. pm-critic’s proactive trigger fires, runs adversarial review, returns P0-P3 findings. The user iterates on the artifact based on findings.

Pattern 2: Slash command -> sub-agent (explicit invocation)

User runs utility-pm-critic, utility-pm-skill-auditor, utility-pm-changelog-curator, or utility-pm-release-conductor. The command invokes the paired sub-agent. Single-turn lifetime for critic and curator; multi-turn for auditor and conductor.

Pattern 3: Sub-agent -> sub-agent (chain composition, 2 levels max)

pm-release-conductor chains to children at specific gates:

• G0 (pre-tag readiness): chains to pm-skill-auditor; receives audit findings via layered Status envelope per D26
• G2 (CHANGELOG prep): chains to pm-changelog-curator; receives draft via layered Status envelope per D26
• G2.5 (commit + re-verify): re-runs G0 chain after committing release-prep edits per D22

Neither child (auditor, curator) chains further. Chain depth = 2 is enforced via agents/_chain-permitted.yaml allowlist + validate-agents-md (D21 HARD FAIL).

Pattern 4: Dispatch skill -> sub-agent or inline execution (cross-client)

Dispatch skill detects runtime and either invokes native sub-agent (Claude Code path) or reads + executes the agent definition inline (non-Claude path). The “reference + execute inline” pattern enables non-Claude clients to consume sub-agent intent without losing functional access.

For pm-release-conductor specifically (most complex sub-agent), the dispatch skill uses an expanded “reference + execute inline” pattern that inlines auditor + curator behaviors at G0 + G2 on non-Claude clients (instead of chaining; chain composition is Claude Code only). This pattern is validated by Phase 2 GATE C sub-spike before shipping the conductor dispatch skill.

Pattern 5: /chain -> orchestrator Mode B -> builder (ad-hoc to durable, v2.26.0)

/chain is a thin front door to pm-workflow-orchestrator Mode B: it parses only the separator-driven chain-expression boundary (grammar in skills/utility-pm-workflow-orchestrator/references/PARSE-CONTRACT.md, Mode B Chain Expression Contract) and hands the steps, flags, and context to the engine, which validates every name pre-flight and owns all run rules. A completed 2+ step chain’s terminal output suggests promotion; utility-pm-workflow-builder turns the chain expression into a staged Workflow Implementation Packet (draft _workflows/ file, draft command, cross-cutting checklist) that a human reviews and promotes. The loop: try a sequence ad hoc, discover it is reusable, make it durable, and from then on run the curated workflow-* command.

Who runs what, pinned (the orchestrator boundary table):

Surface	Persistence	Who executes	Validation source
/chain	Ephemeral; nothing committed	pm-workflow-orchestrator Mode B (native sub-agent on Claude Code; dispatch skill inline branch elsewhere)	Chain-expression contract in PARSE-CONTRACT.md
utility-pm-workflow-orchestrator Mode B direct	Ephemeral	Same engine	Same contract
workflow-* commands	Durable, curated, hand-authored	Main agent reads each skill inline (NOT the orchestrator)	Author judgment + repo validators
utility-pm-workflow-builder	Authors durable files via _staging/	Not an executor; writes a packet only	Same contract, applied at authoring time

The boundary rules carried over unchanged from v2.24.0: the orchestrator never nests a workflow (Category 3 steps surface as MANUAL), never spawns a sub-agent (leaf-inlining for Category 2; no Agent tool; no _chain-permitted.yaml entry), and Mode A / Mode B threading semantics are unchanged apart from the now-named --thread flag for user-declared linear dependency.

Cross-References

• Master plan: docs/internal/release-plans/v2.16.0/plan_v2.16.0.md (D11, D14, D19, D20, D21, D26, D30, D31)
• Subagents integration plan: docs/internal/release-plans/v2.16.0/subagents-integration-plan.md
• Adversarial review user guide: docs/guides/adversarial-review.md (ships in Phase 2)
• Release runbook contributor guide: docs/contributing/release-runbook.md (ships in Phase 5)
• Chain allowlist: agents/_chain-permitted.yaml
• Strategy source: docs/internal/_working/subagents/subagent-strategy_2026-05-07.md