- lfg input resolution: extract-only for Riffrec/video/audio; do not continue
into ce-brainstorm's interactive path, so autonomous runs don't block
- lfg feedback PR template: make narration block-quote and before-frames
conditional on a transcript; screenshot-only runs describe what's shown,
no invented quote
- docs sync: README, docs/skills/lfg.md, and the skill catalog now describe the
widened inputs (description/Riffrec/video/screenshots) and the dogfood +
compound-out stages
Two gaps surfaced by a gate-branch eval of the skill:
- Step 7 (ce-test-browser) had no skip clause, so a docs-only or
cli/api/library/ios-only change still invoked it. Add a skip for diffs that
touch no web-ui surface, mirroring the docs-only skips on steps 3 and 8.
- Step 9's compound-out left it unspecified that ce-compound in headless mode
can also edit an instruction file (AGENTS.md/CONCEPTS.md) for docs/solutions
discoverability; that edit rides into the run's PR via step 10. State it so it
reads as expected output, not stray changes to revert.
Step 1 now passes 'the resolved task'; no body step references $ARGUMENTS, so
the closing 'use the resolved task wherever a step passes $ARGUMENTS' pointed at
nothing. Replace it with a direct hand-off to step 1 and the step-10 PR body.
Re-targets PR #879 onto the root-native repo layout (#967 moved the skill to
skills/lfg/ and the plugin went skills-only). The original PR's cloud-only
.claude/workflows/lfg.js is dropped: it isn't a distributable plugin artifact,
re-implemented the whole pipeline as a second surface that had already drifted
from the skill, and invoked top-level researcher agents that no longer exist
(they're now references/agents inside ce-plan). Instead, the compounding idea
is expressed inside the linear skill itself, with no harness dependency:
- compound-in: planning already fans out ce-plan's institutional-recall
researchers (learnings, repo conventions, git history, best practices) --
made explicit on step 1, no duplicate research step.
- compound-out: a new ce-compound mode:headless step captures the durable,
non-obvious learning into docs/solutions/ so the next run starts ahead;
non-blocking (Documentation skipped is success), shipped via the commit step.
Also brings the warranted parity additions from #879 onto main's step
structure:
- Input resolution: accept a Riffrec bundle, video/audio, or screenshots and
analyze via ce-riffrec-feedback-analysis before planning.
- Reproduce-before-fix for bug fixes, with synthetic/anonymized state.
- A Dogfood step inlining ce-dogfood's diff-scoped behavior (renamed from
ce-dogfood-beta; still disable-model-invocation, so inlined not invoked).
- A fixed PR-body template for feedback-sourced runs.
Dropped the ce-demo-reel demo-capture instruction (no such skill on main).
Co-Authored-By: Kieran Klaassen <kieranklaassen@gmail.com>
Trevin Chow