Repository navigation
Analysis prompt lists a symlinked skill once per harness root (user scope) #202
Description
Version: backpass 0.1.30; the same code is still in 0.1.32.
Setup. A shared skill library with per-harness symlinks (the layout the loadSkills comment calls
"the common way to share skills across harnesses"): 154 skills live in ~/.agents/skills, and
~/.claude/skills/<name> and ~/.codex/skills/<name> are symlinks to them. User-scope defaults load
all three roots (USER_SKILLS_DIRS).
Problem. resolveProjectSkillDirs dedupes roots by realpath, but the roots are distinct real
directories, so each shared skill is loaded three times. renderSkillIndexForAnalysis then prints
every copy. In our case the "Project skills (load-on-trigger)" section was 567 lines and ~187k chars
for 199 unique skills - 88% of each analysis prompt, paid on every transcript.
Suggested fix. Keep loadedCopies / description-budget accounting as is (each copy is billed on
purpose), but dedupe only the index shown to the analysis model, e.g. by realpath of the skill file
or by name + description, first path wins:
export function renderSkillIndexForAnalysis(skills) {
if (!skills.length) return "(this repo has no skills)";
const seen = new Set();
const unique = skills.filter((s) => {
const key = `${s.name}\u0000${s.description || ""}`;
if (seen.has(key)) return false;
seen.add(key);
return true;
});
return unique.map((s) => `- ${s.name} (${s.path}) :: ${s.description || "(no description)"}`).join("\n");
}Measured effect (2 transcripts, backpass analyze --scope user, claude-sonnet-5-5): skill section
187k -> 77k chars, 199 entries each listed once, first-turn tokens per session -43.5% on an equal basis.
Possibly related to #75 (large skill stores inflating prompts), but a different cause: here it is per-skill symlinks across harness roots, not the copied tree.
Activity
kunchenguid commentedon Oct 6, 2026
Speaking as Kun's firstmate.
Thanks for the careful write-up and the measurements. Triaged on latest main
0268201(chore release 0.1.32 #198).Confirmed on main.
resolveProjectSkillDirsinsrc/skills.jsdedupes only whole roots by realpath, and user scope loads three distinct real roots (USER_SKILLS_DIRS=.agents/skills,.claude/skills,.codex/skillsinsrc/config.js).loadSkillsdeliberately follows per-skill symlinks, so one library file reached through three links becomes three entries.renderSkillIndexForAnalysisthen prints every entry, andsrc/analyze.jssends that index with every transcript.loadedCopiesandskillDescriptionTokensintentionally bill each copy, and that should stay as it is. Only the index the analysis model reads needs each file once.This is different from #75 (the workspace copying the whole skills tree). None of the open PRs (#150, #151, #154, #163, #175, #179, #183, #201) touch the skill index.
One note for whoever picks this up: please key the dedupe on the skill file's realpath (the same identity
loadedCopiesalready uses), not on name + description. Two genuinely different files that happen to share a name and description are separate skills, and silently hiding one would trade accuracy for size. A test should fail when the index lists a symlinked library skill more than once, and should still list two distinct files that share a name.Contract-class: restore. The analysis prompt is meant to show the skill index, meaning the skills that exist. Listing one file once per link is a defect on that existing default path. It adds no new surface, write, or default.
VISION per-rule:
coveredBySkillstill sees every distinct skill.Labeling
ready-for-pr.kunchenguid commentedon Oct 20, 2026
Speaking as Kun's firstmate.
Stale-restamp on latest main
c22f2ad(backpass 0.1.37).Still real and unhelped on main:
renderSkillIndexForAnalysisinsrc/skills.js(line 255) still prints every loaded entry with no dedupe, so a library skill reached through per-harness symlinks under.agents/skills,.claude/skills, and.codex/skillsis still listed once per root in every analysis prompt. No open PR changes the skill index (#212 only imports the function). Not opening a competing PR. The earlier note still applies for whoever picks this up: key the dedupe on the skill file's realpath, keep per-copy budget billing as is, and still list two distinct files that share a name.Contract-class: restore (the analysis prompt already promises an index of the skills that exist; listing one file once per link is a defect on that default path, with no new surface, write, or default).
VISION per-rule:
coveredBySkillstill sees every distinct skill).Leaving
ready-for-pr.fix: deduplicate symlinked skills in the analysis index (kunchenguid#231
chore: sync upstream main into fork (#2)