pds

Codebase quality analyst. Use to scan for tech debt, code smells, missing tests, and file findings as GitHub issues.

Documentation specialist. Use when READMEs, changelogs, API docs, or inline documentation need updating after code changes.

Team lead for multi-agent tasks. Use when work needs decomposition, parallel execution, or coordination across agents and worktrees.

Deep codebase exploration. Use when you need thorough analysis of code, patterns, dependencies, or context before planning or implementation.

Code review specialist. Use to review diffs for quality, security, correctness, and best practices.

PDS meta-improvement analyst. Use after completing work to identify improvements to skills, agents, and configuration.

Common behavioral rules inherited by all PDS agents. Not a standalone agent — referenced via `inherits: shared-rules` in agent frontmatter.

Persistent cross-swarm advisor. Walks the ticket alongside workers through Phases 1-6. Enforces PDS whitepaper/philosophy/ethos by citation — advisory only, never blocks. Consult for substance questions (design, trade-offs, principle-checks); route graph questions (dispatch, dependencies, phase state) to the orchestrator.

Merge and test specialist. Use after workers finish to merge branches, run test suites, and verify acceptance criteria.

Implementation specialist. Use for scoped coding tasks in isolated worktrees — writing code, fixing bugs, adding features.

Fixing bugs using a test-first disciplined loop. Use when a bug is reported or discovered and needs a verified fix with minimal blast radius.

Bumping project version and updating changelog atomically. Use when releasing a new version — patch, minor, or major.

Save a checkpoint — bump version, commit, push. Use when work is complete and ready to ship without the full finish protocol.

Contributing changes to PDS skills, agents, or workflow. Use before modifying any PDS artifact to follow the whitepaper-alignment checklist.

Grounding in core development principles. Use when starting significant work, feeling stuck, or needing to reset decision-making clarity.

Testing whether skills produce correct agent behavior. Use after modifying a skill, during periodic review, or when model upgrades may affect skill relevance.

Structural codebase exploration using the codebase-memory-mcp SQLite index when one exists for the current project. Prefer over blind Grep when available.

Export a Claude Code session to human-readable markdown. Use when the user wants to save, review, or share a conversation transcript.

Completing a development branch for merge readiness. Use when implementation and tests pass and the branch needs formal preparation for review and merge.

Interrogating requirements to validate before building. Use before swarm decomposition, design decisions on ambiguous features, or when scope creep risk is high.

Recording, reviewing, and promoting engineering patterns. Use when observing a recurring pattern, during post-swarm analysis, or when an instinct reaches promotion threshold.

Save session state and pause gracefully. Use when stopping work mid-session to preserve context for later resumption.

Address PR review comments systematically. Use when a PR has review feedback that needs to be resolved with code changes.

Focused rebase against a target branch. Use when you need to update a feature branch with upstream changes. Stops on conflicts — does not auto-resolve. Extended mode adds autonomous conflict resolution and test-fix loop.

Launching multi-agent parallel work with the Agentic SDLC. Use when a task benefits from decomposition into parallel subtasks.

Referencing the agent roster, roles, coordination model, and dispatch modes. Use when spawning agents or checking permissions.

Associate every PDS task with a GitHub issue. Orchestrator finds or creates the ticket, posts plan and acceptance criteria as a checkbox list, updates it as work progresses. Use at Phase 1 of every swarm.

Triage insights into actionable GitHub issues across repos. Use after running /insights to convert analysis into tracked work.

Verifying work completeness before declaring done. Use before marking tasks complete, creating PRs, or claiming task status as done.

Terse user-facing register — limited vocabulary, doubled state-transition phrases, no hedging or editorializing. Use in the top-level conversation and for orchestrator user-facing inline status only.

Managing git worktrees for isolated parallel development. Use when creating branches, setting up agent worktrees, or cleaning up after merges.

15 hooks across 13 events