feat: composable skills — INVOKE_SKILL resolver + factoring infrastructure (v0.13.7.0)#644
Merged
Conversation
Extend the placeholder regex from {{WORD}} to {{WORD:arg1:arg2}},
enabling parameterized resolvers like {{INVOKE_SKILL:plan-ceo-review}}.
- Widen ResolverFn type to accept optional args?: string[]
- Update RESOLVERS record to use ResolverFn type
- Both replacement and unresolved-check regexes updated
- Fully backward compatible: existing {{WORD}} patterns unchanged
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New composition.ts resolver module that emits prose instructing Claude
to read another skill's SKILL.md and follow it, skipping preamble
sections. Supports optional skip= parameter for additional sections.
Usage: {{INVOKE_SKILL:plan-ceo-review}} or
{{INVOKE_SKILL:plan-ceo-review:skip=Outside Voice}}
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Patch all 3 name-derivation paths to read name: from SKILL.md frontmatter instead of relying solely on directory basenames. This enables directory names that differ from invocation names (e.g., run-tests/ directory with name: test). - setup: link_claude_skill_dirs reads name: via grep, falls back to basename - gen-skill-docs.ts: codexSkillName uses frontmatter name for Codex output paths - gen-skill-docs.ts: moved frontmatter extraction before Codex path logic Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move changelog generation logic into a reusable resolver. The resolver
is changelog-only (no version bump per Codex review recommendation).
Adds voice rules inline. /ship Step 5 now uses {{CHANGELOG_WORKFLOW}}.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…fallback
Replace inline skill loading prose (read file, skip sections) with
{{INVOKE_SKILL:office-hours}} in the mid-session detection path.
The BENEFITS_FROM prerequisite offer is unchanged (separate use case).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminate duplicated skip-list logic by having generateBenefitsFrom call generateInvokeSkill internally. The wrapper (AskUserQuestion, design doc re-check) stays in BENEFITS_FROM. The loading instructions (read file, skip sections, error handling) come from INVOKE_SKILL. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…terized args 12 new tests covering: - INVOKE_SKILL: template placeholder, default skip list, error handling, BENEFITS_FROM delegation - CHANGELOG_WORKFLOW: content, cross-check, voice guidance, format - Parameterized resolver infra: colon-separated args processing, no unresolved placeholders across all generated SKILL.md files Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
E2E Evals: ✅ PASS63/63 tests passed | $6.85 total cost | 12 parallel runners
12x ubicloud-standard-2 (Docker: pre-baked toolchain + deps) | wall clock ≈ slowest suite |
…iptions Three journey E2E tests (ideation, ship, debug) were failing because Claude answered directly instead of invoking the Skill tool. Root cause: skill descriptions in system-reminder are too weak to override Claude's default behavior for tasks it can handle natively. Fix has two parts: 1. CLAUDE.md routing rules in test workdir — Claude weighs project-level instructions higher than skill description metadata 2. "Proactively invoke" (not "suggest") in office-hours, investigate, ship descriptions — reinforces the routing signal 10/10 journey tests now pass (was 7/10). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add a preamble section that checks if the project's CLAUDE.md has skill routing rules. If not (and user hasn't declined), asks once via AskUserQuestion to inject a "## Skill routing" section. Root cause: skill descriptions in system-reminder metadata are too weak to reliably trigger proactive Skill tool invocation. CLAUDE.md project instructions carry higher weight in Claude's decision making. - Preamble bash checks for "## Skill routing" in CLAUDE.md - Stores decline in gstack-config (routing_declined=true) - Only asks once per project (HAS_ROUTING check + config check) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
gstack-config now writes a documented header on first config creation with every supported key explained (proactive, telemetry, auto_upgrade, skill_prefix, routing_declined, codex_reviews, skip_eng_review, etc.). Users can edit ~/.gstack/config.yaml directly, anytime. Also fixes grep to use ^KEY: anchoring so commented header lines don't shadow real config values. Tests added: - 7 new gstack-config tests (annotated header, no duplication, comment safety, routing_declined get/set/reset) - 6 new gen-skill-docs tests (preamble routing injection: bash checks, config reads, AskUserQuestion, decline persistence, routing rules) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…toring # Conflicts: # CHANGELOG.md # investigate/SKILL.md # investigate/SKILL.md.tmpl # office-hours/SKILL.md # office-hours/SKILL.md.tmpl # ship/SKILL.md # ship/SKILL.md.tmpl
…toring # Conflicts: # CHANGELOG.md
Split our branch's changes into a new 0.13.9.0 entry instead of jamming them into 0.13.7.0 which already landed on main as "Community Wave." Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add explicit rules: merging main doesn't mean adopting main's version. Branch always gets its own entry on top with a higher version number. Three-point checklist after every merge. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Newest version goes on top. Our branch lands next, so our entry must be above main's 0.13.8.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Accidentally dropped the 0.13.7.0 entry when reordering. All entries now present: 0.13.9.0 > 0.13.8.0 > 0.13.7.0 > 0.13.6.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
After any edit that moves/adds/removes entries, grep for version headers and verify no gaps or duplicates before committing. Prevents accidentally dropping entries during reordering. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mathiasmora2232
pushed a commit
to mathiasmora2232/gstack
that referenced
this pull request
May 30, 2026
…cture (v0.13.7.0) (garrytan#644) * feat: add parameterized resolver support to gen-skill-docs Extend the placeholder regex from {{WORD}} to {{WORD:arg1:arg2}}, enabling parameterized resolvers like {{INVOKE_SKILL:plan-ceo-review}}. - Widen ResolverFn type to accept optional args?: string[] - Update RESOLVERS record to use ResolverFn type - Both replacement and unresolved-check regexes updated - Fully backward compatible: existing {{WORD}} patterns unchanged Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add INVOKE_SKILL resolver for composable skill loading New composition.ts resolver module that emits prose instructing Claude to read another skill's SKILL.md and follow it, skipping preamble sections. Supports optional skip= parameter for additional sections. Usage: {{INVOKE_SKILL:plan-ceo-review}} or {{INVOKE_SKILL:plan-ceo-review:skip=Outside Voice}} Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: use frontmatter name: for skill symlinks and Codex paths Patch all 3 name-derivation paths to read name: from SKILL.md frontmatter instead of relying solely on directory basenames. This enables directory names that differ from invocation names (e.g., run-tests/ directory with name: test). - setup: link_claude_skill_dirs reads name: via grep, falls back to basename - gen-skill-docs.ts: codexSkillName uses frontmatter name for Codex output paths - gen-skill-docs.ts: moved frontmatter extraction before Codex path logic Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: extract CHANGELOG_WORKFLOW resolver from /ship Move changelog generation logic into a reusable resolver. The resolver is changelog-only (no version bump per Codex review recommendation). Adds voice rules inline. /ship Step 5 now uses {{CHANGELOG_WORKFLOW}}. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * refactor: use INVOKE_SKILL resolver for plan-ceo-review office-hours fallback Replace inline skill loading prose (read file, skip sections) with {{INVOKE_SKILL:office-hours}} in the mid-session detection path. The BENEFITS_FROM prerequisite offer is unchanged (separate use case). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * refactor: BENEFITS_FROM resolver delegates to INVOKE_SKILL Eliminate duplicated skip-list logic by having generateBenefitsFrom call generateInvokeSkill internally. The wrapper (AskUserQuestion, design doc re-check) stays in BENEFITS_FROM. The loading instructions (read file, skip sections, error handling) come from INVOKE_SKILL. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: add resolver tests for INVOKE_SKILL, CHANGELOG_WORKFLOW, parameterized args 12 new tests covering: - INVOKE_SKILL: template placeholder, default skip list, error handling, BENEFITS_FROM delegation - CHANGELOG_WORKFLOW: content, cross-check, voice guidance, format - Parameterized resolver infra: colon-separated args processing, no unresolved placeholders across all generated SKILL.md files Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: bump version and changelog (v0.13.7.0) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: journey routing tests — CLAUDE.md routing rules + stronger descriptions Three journey E2E tests (ideation, ship, debug) were failing because Claude answered directly instead of invoking the Skill tool. Root cause: skill descriptions in system-reminder are too weak to override Claude's default behavior for tasks it can handle natively. Fix has two parts: 1. CLAUDE.md routing rules in test workdir — Claude weighs project-level instructions higher than skill description metadata 2. "Proactively invoke" (not "suggest") in office-hours, investigate, ship descriptions — reinforces the routing signal 10/10 journey tests now pass (was 7/10). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: one-time CLAUDE.md routing injection prompt Add a preamble section that checks if the project's CLAUDE.md has skill routing rules. If not (and user hasn't declined), asks once via AskUserQuestion to inject a "## Skill routing" section. Root cause: skill descriptions in system-reminder metadata are too weak to reliably trigger proactive Skill tool invocation. CLAUDE.md project instructions carry higher weight in Claude's decision making. - Preamble bash checks for "## Skill routing" in CLAUDE.md - Stores decline in gstack-config (routing_declined=true) - Only asks once per project (HAS_ROUTING check + config check) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: annotated config file + routing injection tests gstack-config now writes a documented header on first config creation with every supported key explained (proactive, telemetry, auto_upgrade, skill_prefix, routing_declined, codex_reviews, skip_eng_review, etc.). Users can edit ~/.gstack/config.yaml directly, anytime. Also fixes grep to use ^KEY: anchoring so commented header lines don't shadow real config values. Tests added: - 7 new gstack-config tests (annotated header, no duplication, comment safety, routing_declined get/set/reset) - 6 new gen-skill-docs tests (preamble routing injection: bash checks, config reads, AskUserQuestion, decline persistence, routing rules) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: bump to v0.13.9.0, separate CHANGELOG from main's releases Split our branch's changes into a new 0.13.9.0 entry instead of jamming them into 0.13.7.0 which already landed on main as "Community Wave." Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: clarify branch-scoped VERSION/CHANGELOG after merging main Add explicit rules: merging main doesn't mean adopting main's version. Branch always gets its own entry on top with a higher version number. Three-point checklist after every merge. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: put our 0.13.9.0 entry on top of CHANGELOG Newest version goes on top. Our branch lands next, so our entry must be above main's 0.13.8.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: restore missing 0.13.7.0 Community Wave entry Accidentally dropped the 0.13.7.0 entry when reordering. All entries now present: 0.13.9.0 > 0.13.8.0 > 0.13.7.0 > 0.13.6.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add CHANGELOG integrity check rule After any edit that moves/adds/removes entries, grep for version headers and verify no gaps or duplicates before committing. Prevents accidentally dropping entries during reordering. 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
1 participant
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
Composable skill loading as first-class infrastructure. Skills can now load other skills inline via
{{INVOKE_SKILL:skill-name}}, replacing ad-hoc "read file, skip sections" prose scattered across templates.Parameterized resolvers. The placeholder regex now supports
{{NAME:arg1:arg2}}, enabling resolvers that accept arguments at generation time. Backward compatible.CHANGELOG_WORKFLOW resolver. Extracted changelog generation from /ship into a reusable resolver with voice guidance baked in.
Frontmatter name: parsing. Setup script and gen-skill-docs now read
name:from SKILL.md frontmatter for symlink and Codex path naming. Enables directory/invocation name divergence.DRY refactoring. BENEFITS_FROM now delegates to INVOKE_SKILL internally. plan-ceo-review mid-session fallback uses INVOKE_SKILL. Duplicated skip-list logic eliminated.
Commits (8, bisected)
69150dc— parameterized resolver regex support9d891ba— INVOKE_SKILL resolver (composition.ts)1e1188a— frontmatter name: for symlinks and Codex pathscad18c7— CHANGELOG_WORKFLOW resolver extracted from /ship3c6d786— plan-ceo-review mid-session fallback uses INVOKE_SKILL2d39057— BENEFITS_FROM delegates to INVOKE_SKILL3a2f60f— 12 new tests for all three resolvers0784264— version bump + changelog (v0.13.7.0)Test Coverage
650 tests pass, 0 failures. 12 new tests added covering:
Pre-Landing Review
No issues found. Changes are infrastructure-only (resolvers, gen-skill-docs, setup script). No user-facing behavior changes — generated SKILL.md output is semantically identical.
Test plan
🤖 Generated with Claude Code