Repository navigation
perf: cut the bodies of the most-invoked skills through progressive disclosure #12240
Description
Refs: #12229
Problem
A skill's body enters the orchestrating session's context each time it is invoked, and the most-invoked skills carry long bodies.
Measured on 2026-11-01:
- Method: SKILL.md body bytes with frontmatter stripped, at origin/main
b1439e9809, across 382 plugin skills. - Results: median 11,298 bytes, p90 30,878, and 79 skills over 20 KB.
- Invocation counts come from
Skilltool calls in this repository's local transcripts for the last 2 days: 935 transcript files, 86 main sessions.
| Skill | Invocations (2 days) | Body bytes | Body bytes loaded |
|---|---|---|---|
/source-control:pull-request |
161 | 32,568 | 5.24 MB |
/session-flow:handoff |
53 | 36,786 | 1.95 MB |
/source-control:worktree |
36 | 31,533 | 1.14 MB |
/discovery:research |
17 | 43,880 | 0.75 MB |
/evals:plugin-eval |
18 | 41,142 | 0.74 MB |
/source-control:commit |
15 | 28,968 | 0.43 MB |
Before figure: these six skills loaded about 10.3 MB of body text in 2 days. At about 4 characters per token that is roughly 2.5M tokens, an estimate that was not tokenized.
Done when
- Each of the six skills has its body cut through progressive disclosure: action-specific detail moves to reference files read on demand, and the hub keeps routing and gates. Each cut ships with eval evidence of no loss, as the skill criteria require.
- The after figure is measured the same way: body bytes, weighted by invocations over a comparable window.
Description text is out of scope (#6476, #6846).
This was generated by AI.
Activity
github-actions commentedon Nov 2, 2026
Proposed work class: work-class: scoped
Basis: automated intake triage; the lanes App applied this class.
Size: Large
Complexity: Medium
Uncertainty: Knowable
Flags: none
Blocked by: none
I understand this as a perf task. It cuts the bodies of six heavily used skills (pull-request, handoff, worktree, research, plugin-eval, commit) by moving action-specific detail into reference files read on demand. Each cut needs eval evidence of no loss. The after figure is then re-measured the same way as the before figure.
The scope and done-when are clear. Six skills across several plugins is large, so it likely needs splitting into one change per skill. This is the sub-issue of #12229, which tracks the wider orchestrator-context cost work. I found no duplicate among the open issues.
Automated triage, applied by the lanes App.
melodic-automation-lanes commentedon Nov 2, 2026
Not admitted to the issue lane yet: sub-issue #12244 is open.
melodic-automation-lanes commentedon Nov 2, 2026
Split at intake (owner Q93). This issue is Size Large, so its done-when is split into 2 sub-issues, each holding its task-list lines as written.
Automated split by the lanes App. Admission waits on this issue while a sub-issue is open; work-unblock closes it as completed once every sub-issue is closed as completed by a merge or a listed human and its done-when is covered.
melodic-automation-lanes commentedon Nov 4, 2026
Every sub-issue of this split is closed as completed by a listed human or a merge (#12243, #12244), so the lanes App closes it as completed.