[lukemaj]

What: The loader Skill now installs on all four hosts and loads before the first edit on authoring, audit and tool-description requests; the benchmark and discovery output are recorded as a case study.
Why: plugin.json gave author as a string, so Claude Code refused the plugin and Grok Build installed it with no Skills; separately, the description never named tool descriptions (0/10 loads on that prompt).
So what: Needs independent review, then the user's merge. Behavioral verification in real projects stays with the user.

Closes #17
Refs #7

Changes

• .claude-plugin/plugin.json: "author": { "name": "lukaszmaj" } (the only manifest change; the docs need no skills field).
• Loader description: one before-clause, "before writing or changing an MCP or agent tool description".
• case-studies/2026-09-30-loader-trigger-benchmark.md plus evidence manifest, and case-studies/evidence/2026-09-30-loader-trigger-benchmark/ (84 records, summary.md, discovery.txt, trigger_test.py, discover.py).
• tests/check-loader-benchmark-records.test.sh: summary.md must regenerate byte for byte from records.jsonl, and the scorer self-test must pass.
• One sentence in docs/08-packaging-as-plugin.md, README and docs.json entries, CHANGELOG [Unreleased] lines, regenerated public/llms.txt.

Results (from summary.md)

Host, model	author	audit	tool: before / after	negative clean
Claude Code, Opus 5.5 medium	5/5	5/5	0/5 / 4/5 (5/5 loaded, 1 read then no edit)	3/3, 3/3
Codex, gpt-6-astra low	5/5	5/5	0/5 / 5/5	3/3, 3/3
Grok Build, grok-4.7 medium	3/3	3/3	3/3 / not re-run	2/2
OpenCode, Muse 1.3	3/3	3/3	2/3 / not re-run	2/2

Control (copy present, not installed): Codex 0/5; Claude Code 1/5, by browsing the --add-dir copy. Arm refs f8db18e and eec057d predate the rebase and stay on branch evidence/17-loader-benchmark-arms.

Proof

• Discovery per host, main vs fix, no model call: discovery.txt (Claude Code author: Invalid input / loaded; Grok Skills (0) / Skills (1); Codex and OpenCode listed both times).
• Docs cited with dates: Claude Code plugin manifest reference and Codex Build skills, read 2026-09-30.
• Confinement proven with real denials before the batch on every host (probe file under the real home: read and write denied; loader copy readable, repository writable); guard, cleanup and scorer self-checks passed; no guard abort in 84 runs.
• npm ci && npm test on Node 22.23.1 at 8a64ed6: exit 0 (includes the new records test and the Gate version anchors and CHANGELOG entries in npm test #22 version and changelog gate).
• The new test rejects a scorer that counts refused reads (expected loaded-only, got fired) and an edited record.

Deviations from the ticket

• Hosts: the ticket scoped Claude Code and Codex; the planner added Grok Build and OpenCode at about 60 runs, so those two ran 3 per cell with no control or re-measure. Total 84 runs (68 batch + 16 re-measure).
• The tool-description case was added at the planner's request (from Tool descriptions are activation contracts; deferred loading hides them #4).
• Control: 5 pooled runs per host (2, 2, 1 per prompt), below 5 per prompt and arm; the case study limits it to a floor for loads without installation and draws no per-prompt comparison.
• No version bump: the release step belongs to the user per AGENTS.md; entries are under [Unreleased].

Elon record

• Requirements and who asked: Verify the loader Skill is discoverable and triggers on authoring and audit requests #17 (user, via Move TODO.md, the drift-prevention draft and the continuous-learning handoff into Issues #11): verify discovery against current docs, benchmark per trigger-benchmarks, record a case study, fix the smallest mechanism on a miss. Planner: four hosts and models, about 60 runs, a tool-description prompt, remeasure after a description change.
• Deleted: main-arm model runs (discovery shows the plugin does not load on Claude Code and Grok, so those runs would score 0 by construction); a per-row sweep of all loader routes; the draft's skills: field fix (docs show it is unnecessary); control, re-measure and 5-run cells on Grok and OpenCode (outside the ticket's hosts and the run budget).
• Bottleneck: discovery. The Skill never loaded on two hosts because of one manifest field, so no wording could matter until that was fixed.
• Checked myself: ran claude plugin validate, grok plugin validate, grok inspect, codex debug prompt-input, opencode debug skill on both refs; read the call list of every miss and of the control hit; recomputed every number in the case study from records.jsonl; verified manifest hashes.

🤖 Generated with Claude Code