mirror of
https://github.com/adobe/skills.git
synced 2026-09-19 08:06:09 +08:00
Himani Chauhan
4de24dec1d
* perf(migration): collapse runbook generation to a one-command flow; fix path leak Runbook generation was taking ~19 min not because of compute (all scanners + greps run sub-second) but because the Step-0 workflow forced ~21 model round-trips. This removes the main drivers and fixes the generator bug that triggered a verify->fix->regenerate loop. - One-command CLI as the canonical Step-0 path: add `--bpa-json` (pre-fetched CAM/MCP findings, no in-process mcpFetcher bridge) and `--llm-findings` (Tier-4 results merged + re-rendered + re-cached in one pass). The agent no longer reads the generator source or hand-mutates `result.gathered`. - Union the local analyzer with BPA instead of short-circuiting: BPA stays authoritative, the analyzer fills source-only cascade findings BPA's scan scope missed (deduped by class name, rendered `BPA / CAM + analyzer`). Removes the manual "BPA clean vs real source" reconciliation and is a correctness win. - Fix relPath: a finding whose path IS the workspace root (e.g. a Dispatcher config at the project root) rendered as the absolute path; now collapses to '.' so the runbook never leaks the local filesystem layout. - SKILL.md Step 0 rewritten around the single command; programmatic API demoted to an "Advanced" note. scripts/README documents the new flags. - Tests: relPath root-collapse, classKey dedup, sourceLabel unions, preFetchedBpa (no-bridge), mergeLlmFindings, end-to-end llmByPattern, and an analyzer-union+dedup test (skips cleanly without a JDK). 152 tests, 0 fail. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(migration): close BPA-incomplete gaps, add Felix SCR scheduler + AMS mode Follow-up to the one-command runbook flow: three real gaps surfaced running it against a live project where a BPA/CAM source was attached, plus a mislabelled dispatcher mode. - bpaIncomplete: a BPA source present but structurally unable to report a pattern no longer marks it clean. templateModernization (CAM can't see templates) now falls through to the local content-scan; guavaCache (bpa-only; CAM filters the project's own bundles) falls through to the LLM tier instead of silent-clean. - Scheduler detector: recognise the legacy Felix SCR form (org.apache.felix.scr.annotations @Service(Runnable.class) + @Property/@Properties with a scheduler.* property), which the OSGi-DS-only detector missed. Adds FelixScrScheduler (positive) + FelixPlainRunnable (negative) fixtures. - Dispatcher detectMode: a monolithic layout WITH AMS markers (ams_* / *_farm.any / conf.d/whitelists) is now labelled `ams` instead of a misleading `flexible`. Same on-premise executor route as flexible — label only; added to CONVERTIBLE. - Docs: SKILL.md + references/dispatcher/context.md mode taxonomy updated. - Tests: migration JS suite 158 (156 pass, 2 skipped, 0 fail); analyzer harness 115 pass, 0 fail; both skills validate. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Himani Chauhan <hichauha+adobe@adobe.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>