fix(subagent): lean substrate — no skills by default + opt-in skills param (#32) - #42
Merged
Merged
Conversation
…param (#32) Root cause (#32): buildChildLoader's skillsOverride loaded ALL installed skills (~42 → ~570K tokens, ~59% of a 976K context window from turn 1) when an agent declared no 'skills' field. Both builtin agents (general-purpose, general- purpose-cc) declare none → every dispatch paid the full substrate cost before doing any work. Fix (default (a) — lean, caller opts in): - src/engine/child-loader.ts: extract resolveChildSkills() (pure, testable); the no-skills branch now returns [] (not cur.skills). An agent that declares specific skills still loads only those (truthy branch unchanged). - src/tools/subagent.ts: add a 'skills' param (Type.Array(Type.String())) — callers opt in to specific skills. Threaded as skillsOverride to the direct spawn path + (params.skills ?? o.skills) to the lifecycle spawn adapter so a caller can override a lifecycle phase's bundle. Blast radius (challenge-step verified): - general-purpose / general-purpose-cc direct dispatches → now load 0 skills (lean substrate). Callers pass 'skills' to opt in. - default lifecycle phases all declare explicit skill bundles → unaffected (truthy branch, filters to declared skills). - verify.ts gate reviewer passed skills:[] → was also loading all 42 (the old else-branch); now correctly loads 0. Bonus lean-up, no regression. Tests (+6): child-loader.test.mts — resolveChildSkills for no-skills → [], specific-skills filter, explicit-empty → [], unknown-names dropped. subagent- tool.test.mts — skills param threads to child agent.skills; absent → undefined. 621/621 pass, typecheck clean. Substrate reduction to be confirmed via the release-gate smoke (real pi dispatch) at tag time; the other #32 directions (memory-hydration trim, prompt compaction, widget UX) remain tracked in #32.
…de, EOF newlines) Review findings addressed (APPROVE w/ nits): - src/tools/subagent.ts: lifecycle spawn adapter switched from 'params.skills ?? o.skills' (REPLACE — a caller passing skills:["tdd"] with lifecycle:"default" would strip 'brainstorming' from the brainstorm phase) to mergeLifecycleSkills(o.skills, params.skills) — ADDITIVE + deduped. The phase's designed skills always load; a caller can add extras but cannot strip phase skills. Extracted mergeLifecycleSkills() as a pure testable helper. - src/tools/subagent.ts: clarified the skills param description — direct dispatch replaces frontmatter (pass [] for zero); lifecycle dispatch is additive. Fixes the description-accuracy nit. - src/engine/child-loader.ts + src/tools/subagent.ts: added EOF newlines (AGENTS.md convention; the two source files were missing them). Locked in the [] override behavior (a reviewer worried skills:[] was silently ignored — it isn't: [] is truthy, the clone happens, resolveChildSkills returns []). Added a test for it on an agent WITH frontmatter skills. Tests (+2): skills:[] overrides frontmatter to zero (direct); mergeLifecycleSkills additive + deduped (5 cases).
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.
What
Root-cause fix for #32:
buildChildLoader'sskillsOverrideloaded ALL installed skills (~42 → ~570K tokens, ~59% of a 976K context window from turn 1) when an agent declared noskillsfield. Both builtin agents (general-purpose,general-purpose-cc) declare none → every dispatch paid the full substrate cost before doing any work.Default chosen per RECTOR: (a) none — lean substrate, caller opts in.
Changes
src/engine/child-loader.ts— extractedresolveChildSkills()(pure, testable); the no-skills branch now returns[](notcur.skills). An agent that declares specific skills still loads only those (truthy branch unchanged).src/tools/subagent.ts— added askillsparam (Type.Array(Type.String())); callers opt in to specific skills. Threaded asskillsOverrideto the direct spawn path + (params.skills ?? o.skills) to the lifecycle spawn adapter so a caller can override a lifecycle phase's bundle. Added a promptGuidelines note.Blast radius (challenge-step verified)
general-purpose/general-purpose-ccdirect dispatchesskillsdefaultlifecycle phases (all declare explicit skill bundles)verify.tsgate reviewer (passedskills:[])No silent skill loss: lifecycle phases declare skills explicitly, so they hit the truthy branch.
Tests (+6)
test/child-loader.test.mts—resolveChildSkills: no-skills →[]; specific-skills → filtered; explicit-empty →[]; unknown names dropped gracefully.test/subagent-tool.test.mts—skillsparam threads to the clonedchild agent.skills; absent → staysundefined(lean).Verification
pnpm typecheck— cleanpnpm test:run— 621/621 pass (was 615; +6 new)Not in this PR
The other #32 directions — memory-hydration trim, system-prompt compaction, widget work-progress UX (overlaps #23) — remain tracked in #32. The actual substrate-token reduction will be confirmed via the release-gate smoke (real pi dispatch) at tag time; this PR proves the skill-loading logic only.
Addresses #32 (the skill-loading root cause). The remaining #32 optimization directions stay open.