feat: multi-agent support — gstack works on Codex, Gemini CLI, and Cursor (v0.9.0)#226
Merged
feat: multi-agent support — gstack works on Codex, Gemini CLI, and Cursor (v0.9.0)#226
Conversation
Refactor gen-skill-docs.ts for multi-agent support: - Add Host type, HostPaths interface, HOST_PATHS config - Decompose generatePreamble() into 7 composable sub-functions - Replace all hardcoded .claude/skills/gstack paths with ctx.paths - Replace static findTemplates() list with dynamic filesystem scan - Add --host codex|agents flag (aliases, same output) - Add processTemplate host routing to .agents/skills/gstack-*/ - Add codexSkillName() with double-prefix prevention - Add transformFrontmatter() — keeps only name + description for Codex - Add extractHookSafetyProse() — converts hooks to inline advisory - Add body text path rewriting for remaining hardcoded paths - Exclude /codex skill from Codex generation (self-referential) Claude output is unchanged (verified via --dry-run). SKILL.md is an open standard: .agents/skills/ works on Codex, Gemini CLI, and Cursor. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Generated 21 skill files for the open SKILL.md standard: - Output: .agents/skills/gstack-*/SKILL.md (one per skill) - Frontmatter: name + description only (no allowed-tools/version) - No .claude/skills/ paths in any generated file - /codex skill excluded (Claude wrapper, self-referential on Codex) - Hook skills (careful/freeze/guard) get inline safety prose - Build script generates both hosts: bun run build Supported agents (all read .agents/skills/): - Codex CLI - Gemini CLI - Cursor Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- setup: add --host codex|claude|auto flag, install to ~/.codex/skills/ when targeting Codex, auto-detect installed agents - find-browse: priority chain .codex > .agents > .claude (both workspace-local and global) - dev-setup/teardown: create .agents/skills/gstack symlinks for dev mode Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tests (28 new): - Codex output path routing, frontmatter validation (name+description only) - No .claude/skills/ path leaks in Codex output (regression guard) - /codex skill exclusion, hook→prose conversion, multiline YAML - --host agents alias, dynamic template discovery - Codex skill validation + $B command validation - find-browse priority chain verification - Replace static ALL_SKILLS list with dynamic filesystem scan CI: - Add Codex freshness check to skill-docs workflow Docs: - AGENTS.md: Codex-facing project instructions - README: multi-agent installation section - CONTRIBUTING: dual-host development workflow - CHANGELOG: v0.9.0 multi-agent support entry Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New test infrastructure: - CodexSessionRunner: spawns codex exec, parses JSONL stream, returns structured results (output, reasoning, toolCalls, tokens) - JSONL parser ported from Python (codex/SKILL.md.tmpl) to TypeScript - Temp HOME skill installation for Codex discovery testing E2E tests (gated behind EVALS=1 + codex + OPENAI_API_KEY): - codex-discover-skill: installs skill, verifies Codex finds it - codex-review-findings: runs gstack-review via Codex, validates output Integrates with existing eval infrastructure: - Diff-based test selection via touchfiles - Eval persistence via EvalCollector - bun run test:codex / test:codex:all convenience scripts Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two bugs found by Codex adversarial review: 1. Sidecar path mismatch: generated Codex skills referenced .agents/skills/gstack-review/checklist.md but setup creates sidecars at .agents/skills/gstack/review/. Fixed path rewriter to emit .agents/skills/gstack/review/ (matching setup layout). 2. Setup installed Claude-format source dirs for Codex global install instead of the generated Codex-format skills. Split link_skill_dirs into link_claude_skill_dirs (source dirs for Claude) and link_codex_skill_dirs (generated .agents/skills/ gstack-* dirs for Codex). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
17 new tests covering: - Sidecar path rewriting: .claude/skills/review → .agents/skills/gstack/review/ (catches the bug where checklist.md was unreachable at gstack-review/) - All 4 path rewrite rules tested individually across all skills - Greptile triage sidecar path correctness - Ship skill sidecar paths for pre-landing review - Claude output regression guard: zero Codex paths in any Claude skill - Setup script validation: separate link functions for Claude vs Codex, link_codex_skill_dirs reads from .agents/skills/, create_agents_sidecar links runtime assets (bin, browse, review, qa) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Resolved conflicts in VERSION (keep 0.9.0), CHANGELOG.md (prepend 0.9.0 entry above 0.8.5), and gen-skill-docs.ts (keep dynamic scan over static list). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove stale gstack-debug, add gstack-investigate, regenerate all Codex skills to pick up changes merged from main (investigate rename, platform-agnostic templates, review helpers). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Resolved conflicts: - VERSION: keep 0.9.0 (ours supersedes 0.8.6) - CHANGELOG: prepend 0.9.0 above new 0.8.6 entry - gen-skill-docs.ts: integrate telemetry (pending finalize, prompt) into decomposed sub-functions with host-aware paths - gen-skill-docs.test.ts: keep both Codex generation and telemetry tests Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove OPENAI_API_KEY gate from test prerequisites - Copy real ~/.codex/ auth config into temp HOME so codex can authenticate - Increase review test timeout to 540s (codex does thorough 60+ tool call reviews) - Document in CLAUDE.md that Codex uses its own auth config Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Great it! Big thanks! |
rapidstartup
pushed a commit
to rapidstartup/gstack
that referenced
this pull request
Mar 29, 2026
…rsor (v0.9.0) (garrytan#226) * refactor: host-aware gen-skill-docs + --host codex generation Refactor gen-skill-docs.ts for multi-agent support: - Add Host type, HostPaths interface, HOST_PATHS config - Decompose generatePreamble() into 7 composable sub-functions - Replace all hardcoded .claude/skills/gstack paths with ctx.paths - Replace static findTemplates() list with dynamic filesystem scan - Add --host codex|agents flag (aliases, same output) - Add processTemplate host routing to .agents/skills/gstack-*/ - Add codexSkillName() with double-prefix prevention - Add transformFrontmatter() — keeps only name + description for Codex - Add extractHookSafetyProse() — converts hooks to inline advisory - Add body text path rewriting for remaining hardcoded paths - Exclude /codex skill from Codex generation (self-referential) Claude output is unchanged (verified via --dry-run). SKILL.md is an open standard: .agents/skills/ works on Codex, Gemini CLI, and Cursor. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: generate Codex/Gemini/Cursor skills into .agents/skills/ Generated 21 skill files for the open SKILL.md standard: - Output: .agents/skills/gstack-*/SKILL.md (one per skill) - Frontmatter: name + description only (no allowed-tools/version) - No .claude/skills/ paths in any generated file - /codex skill excluded (Claude wrapper, self-referential on Codex) - Hook skills (careful/freeze/guard) get inline safety prose - Build script generates both hosts: bun run build Supported agents (all read .agents/skills/): - Codex CLI - Gemini CLI - Cursor Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: dual-host setup + find-browse for Codex/Gemini/Cursor - setup: add --host codex|claude|auto flag, install to ~/.codex/skills/ when targeting Codex, auto-detect installed agents - find-browse: priority chain .codex > .agents > .claude (both workspace-local and global) - dev-setup/teardown: create .agents/skills/gstack symlinks for dev mode Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: Codex generation tests + CI + docs for multi-agent support Tests (28 new): - Codex output path routing, frontmatter validation (name+description only) - No .claude/skills/ path leaks in Codex output (regression guard) - /codex skill exclusion, hook→prose conversion, multiline YAML - --host agents alias, dynamic template discovery - Codex skill validation + $B command validation - find-browse priority chain verification - Replace static ALL_SKILLS list with dynamic filesystem scan CI: - Add Codex freshness check to skill-docs workflow Docs: - AGENTS.md: Codex-facing project instructions - README: multi-agent installation section - CONTRIBUTING: dual-host development workflow - CHANGELOG: v0.9.0 multi-agent support entry Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: Codex E2E test harness — verify skills work on Codex CLI New test infrastructure: - CodexSessionRunner: spawns codex exec, parses JSONL stream, returns structured results (output, reasoning, toolCalls, tokens) - JSONL parser ported from Python (codex/SKILL.md.tmpl) to TypeScript - Temp HOME skill installation for Codex discovery testing E2E tests (gated behind EVALS=1 + codex + OPENAI_API_KEY): - codex-discover-skill: installs skill, verifies Codex finds it - codex-review-findings: runs gstack-review via Codex, validates output Integrates with existing eval infrastructure: - Diff-based test selection via touchfiles - Eval persistence via EvalCollector - bun run test:codex / test:codex:all convenience scripts Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: bump VERSION to 0.9.0 to match CHANGELOG Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: Codex sidecar paths + setup installs generated skills Two bugs found by Codex adversarial review: 1. Sidecar path mismatch: generated Codex skills referenced .agents/skills/gstack-review/checklist.md but setup creates sidecars at .agents/skills/gstack/review/. Fixed path rewriter to emit .agents/skills/gstack/review/ (matching setup layout). 2. Setup installed Claude-format source dirs for Codex global install instead of the generated Codex-format skills. Split link_skill_dirs into link_claude_skill_dirs (source dirs for Claude) and link_codex_skill_dirs (generated .agents/skills/ gstack-* dirs for Codex). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: comprehensive Codex path rewriting + setup install tests 17 new tests covering: - Sidecar path rewriting: .claude/skills/review → .agents/skills/gstack/review/ (catches the bug where checklist.md was unreachable at gstack-review/) - All 4 path rewrite rules tested individually across all skills - Greptile triage sidecar path correctness - Ship skill sidecar paths for pre-landing review - Claude output regression guard: zero Codex paths in any Claude skill - Setup script validation: separate link functions for Claude vs Codex, link_codex_skill_dirs reads from .agents/skills/, create_agents_sidecar links runtime assets (bin, browse, review, qa) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: regenerate Codex skills after investigate rename merge Remove stale gstack-debug, add gstack-investigate, regenerate all Codex skills to pick up changes merged from main (investigate rename, platform-agnostic templates, review helpers). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: Codex E2E uses ~/.codex/ auth, not OPENAI_API_KEY - Remove OPENAI_API_KEY gate from test prerequisites - Copy real ~/.codex/ auth config into temp HOME so codex can authenticate - Increase review test timeout to 540s (codex does thorough 60+ tool call reviews) - Document in CLAUDE.md that Codex uses its own auth config Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
gstack now works on any AI agent that supports the open SKILL.md standard. One install, four agents.
.agents/skills/for Codex CLI, Gemini CLI, and Cursor--host codexflag for gen-skill-docs — generates adapted output with Codex paths, stripped frontmatter, inline safety prose for hook skills./setup --host codex|autoinstalls generated skills to~/.codex/skills/with proper sidecar file layoutfind-browsepriority chain checks.codex>.agents>.claudefor binary resolutionCodexSessionRunnerspawnscodex exec, parses JSONL output, validates skill discoveryTest Coverage
Pre-Landing Review
No issues found. Build tooling and test harnesses only — no SQL, user input, or LLM trust boundaries.
Codex Adversarial Review
Codex found 5 issues (3 P1, 2 P2). 2 were real bugs (sidecar paths, setup install format) — both fixed. 1 was not our change (URL validation from main). 2 were valid-but-by-design (Gemini/Cursor use repo-local discovery, E2E tests validate content not install plumbing).
TODOS
No TODO items completed in this PR.
Test plan
.claude/skills/paths in any Codex output🤖 Generated with Claude Code