Agent Claude Code
The cast of subagents the Ralph loop dispatches to build one backlog issue per tick. The conductor is the main Ralph session (.claude/commands/ralph-tick.md → scripts/ralph/PROMPT.md); it spawns every agent below via the Agent tool. The chief-architect is the strategic brain (plans + a dispatch list); the specialists…
Agent Claude Code
Strategic brain of a Ralph tick. Select to architect a single backlog issue: read the issue + house rules, decide the design approach, and produce an ordered dispatch plan naming which specialists the conductor should invoke (test, implementation, security, performance, documentation, dependency). Plans; never writes…
Agent Claude Code
Gate 2.5 — the pre-push self-review. Routes a working-tree diff to the relevant specialist reviewers (test, implementation, security, performance, documentation, dependency) and returns one consolidated, deduplicated, severity-ranked findings report so the conductor fixes slop before it ever reaches CI or the PR.
Agent Claude Code
Read-only review of dependency changes — version pinning, lockfile integrity, transitive conflicts, license compatibility, dev/prod separation. Select when a change touches pyproject.toml or uv.lock (creek-tools/ or crawdad/). Reports findings; does not edit code.
Agent Claude Code
Writes and updates documentation for a change — Python docstrings (interrogate ≥95%), README/module docs, and ADRs. Select when the chief-architect flags a docs gap (new public API or changed behavior), and as the documentation-dimension reviewer. Docs must match the implementation exactly.
Agent Claude Code
Gate 1 GREEN + Refactor — writes the production code that makes the failing tests pass at threshold quality, then refactors. Select for implementing a planned change (pipeline stages, CLI commands, MCP tools — per shared/house-rules.md's stack) and as the correctness/maintainability reviewer. The core code-quality…
Agent Claude Code
Profiles and optimizes performance-sensitive code — O(n²) pairwise link passes, embedding/LLM API call volume, memory at 35k-fragment scale, algorithmic complexity. Select when the chief-architect flags a performance risk, and as the performance-dimension reviewer. Measure first; never trade correctness for speed.
Agent Claude Code
One parallel worker of the Ralph fleet. Select to drive a SINGLE assigned backlog issue through Gates 1–2.5 and open its PR, working entirely inside a dedicated git worktree so it never collides with sibling workers. It is the per-issue conductor (spawns chief-architect + specialists per scripts/ralph/PROMPT.md); it…
Agent Claude Code
Hardens and audits security-sensitive code — MCP transport auth (bearer tokens), secrets, input validation on ingest parsers, vault path traversal, subprocess use, the redaction pipeline, file/network I/O. Select when the chief-architect flags a security risk, and as the security-dimension reviewer. Applies the repo…
Agent Claude Code
Gate 1 RED — writes the failing tests that specify a behavior before it exists, per the chief-architect's test strategy. Select for TDD test authoring and as the test-dimension reviewer (coverage, assertions, edge/error cases). Pytest with the repo's conftest.py fixtures and integration/e2e markers, per…
Command Claude Code
One tick of the local Ralph loop. Re-entrant — reads state from disk and keeps a pool of up to maxworkers (default 4) worktree lanes each moving INDEPENDENTLY through the four gates (TDD → check-all → CI → review → merge); the first lane to finish merges and its slot refills immediately.
Hook Claude Code
Runs when a session starts, executing session-start.sh. From Geoffe-Ga/Creek-Vault.
Settings file Claude Code
Agent settings declaring 1 hook event (SessionStart).
Skill Claude CodeCodex
Iterate on Claude PR review feedback intelligently and merge when ready. Use when the user asks to "address feedback", "respond to Claude's review", "iterate on the PR", "fix review comments", or "merge if Claude said LGTM". The Claude reviewer publishes a top-level PR comment via GitHub Action ending in a Verdict…
Skill Claude CodeCodex
Subscribe to GitHub PR activity and wait — without polling — for the Claude reviewer's verdict comment on the current HEAD. Use when a PR has just been pushed and you need to know the verdict (LGTM / CHANGESREQUESTED / COMMENTS) before merging, addressing feedback, or declaring work done. Wraps…
Skill Claude CodeCodex
Systematic GitHub backlog maintenance: review merged PRs, close resolved issues, identify gaps, and create missing issues. Use when asked to groom the backlog, clean up the issue tracker, or review recent work. Do NOT use for PR code review (use comprehensive-pr-review).
Skill Claude CodeCodex
Debug CI test failures on pull requests with structured protocol. Use when CI fails on your PR, tests pass locally but fail in CI, or you're tempted to say "pre-existing issue". Covers state comparison, error reading, local reproduction, and common root causes. Do NOT use for local-only test failures or feature…
Skill Claude CodeCodex
Structured 10-section PR review covering security, quality, testing, and documentation. Use when reviewing pull requests, evaluating code changes, or doing code review. Produces verdicts with specific references. Do NOT use for backlog grooming or issue triage (use backlog-grooming), reviewing your own changes before…
Skill Claude CodeCodex
Safe and efficient concurrent code patterns across languages. Use when implementing async operations, parallel processing, thread pools, or managing shared state. Covers Python asyncio, TypeScript Promises, Go goroutines, and Rust Tokio. Do NOT use for general performance optimization without concurrency.
Skill Claude CodeCodex
Launch the CrawDad Discord bot against a Creek vault. Use when the user asks to "launch CrawDad", "start the bot", "run CrawDad", "fire up the Discord bot", or otherwise bring the Discord-side interface online for a vault. ALWAYS ask which vault directory to run against first. Handles every launch gotcha (absolute…
Skill Claude CodeCodex
Gentle, hand-held, end-to-end walkthrough of the Creek Vault + CrawDad system, voiced by CrawDad. Use when the user wants to be walked through using Creek — "walk me through", "how do I use this", "get me started", "show me the proper way", "I don't want to read the docs", or any request for guided onboarding. Claude…
Skill Claude CodeCodex
Remediate dependency vulnerability scanner failures by verifying live package registry data and upgrading instead of suppressing. Use when an SCA / CVE tool fails or files an alert: npm audit, pnpm audit, yarn audit, pip-audit, safety, Snyk, Dependabot, Trivy, Grype, OSV-Scanner, audit-ci, cargo audit, govulncheck…
Skill Claude CodeCodex
Audit the codebase for sloppy, amateurish, or AI-generated low-quality code — bugs, code smells, dead/stubbed/orphaned code, duplication, poor architecture, needless verbosity, comment slop, type-safety erosion, and weak tests — then file corroborated findings as Ralph-ready GitHub issues (and decomposed epics for big…
Skill Claude CodeCodex
Write clear, comprehensive documentation with language-specific patterns. Use when writing docstrings, module docs, README files, API references, or architecture decision records. Covers Python, TypeScript, Go, and Rust documentation conventions.