mirror of
https://github.com/adobe/skills.git
synced 2026-09-19 08:06:09 +08:00
Abhishek Garg
620c888761
The tessl-review judge scored content at 48% (overall 78%, below the 80% threshold) on the prior commit, complaining specifically that SKILL.md duplicated ~40% of content already in reference files (threat model, consent flow, full rules checklist, helper op list, implementation-level details like O_NOFOLLOW / TOCTOU / fcntl.flock / /proc/self/fd/N). Trim is information-preserving — every claim that left SKILL.md now lives in a reference, with SKILL.md pointing at it: - New references/threat-model.md (~26 lines) holds the defended trust boundaries table and the explicit out-of-scope items (NL prompt injection, helper supply-chain, adversarial Windows). SKILL.md keeps a 5-line pointer. - references/collision-rules.md gains a "Root CLAUDE.md consent prompt" section (~50 lines) covering state detection (missing / skill-owned / human-curated), decision flow, persistence under decision: claude-md, and CI suppression. SKILL.md keeps a 9-line pointer. - Deterministic helper section: collapsed the op list (already exhaustive in helpers.md) into a 6-line pointer; dropped the implementation-level jargon (O_NOFOLLOW, /proc/self/fd/N, F_GETPATH). - Concurrency / idempotency / modes: condensed from 9 lines to 3 bullets pointing at upgrade-and-migration.md § 1 for byte-exact marker rules. - Rules checklist: 16 verbose bullets → 16 short bullets, each linking to where the rule is authoritative instead of restating it inline. - IDE selection suppression rules: 3 paragraphs collapsed into one with a pointer to output-format.md § 1.1. - Hard-guarantee tail prose: 3 paragraphs collapsed to 2. SKILL.md: 389 → 297 lines (24% reduction). 84/84 helper tests still pass. No information lost — every claim still has a single authoritative home in the references.