feat(lfg): fold compounding into the skill and bring it to parity

[kieranklaassen]

Summary

/lfg was linear and amnesiac: every run started from zero and ended teaching the next one nothing. This PR folds the compounding loop directly into the linear /lfg skill and brings the skill up to parity on the capabilities it was missing — entirely inside the skill, with no harness-specific dependency.

Re-scoped from the original PR. This branch was rebased onto the current root-native repo layout (the skill moved to skills/lfg/ and the plugin went skills-only in #967). The original .claude/workflows/lfg.js cloud workflow has been dropped — it was not a distributable plugin artifact, re-implemented the whole pipeline as a second surface that had already drifted from the skill, and directly invoked top-level researcher agents that no longer exist (they're now references/agents/ inside ce-plan). The valuable idea — compounding — is now expressed in the skill itself.

What's new

Compounding, in the skill:

• Compound IN — planning already fans out ce-plan's institutional-recall researchers (prior learnings in docs/solutions/, repo conventions, git history, best practices), so a run starts from accumulated knowledge instead of a blank slate. Made explicit on step 1; no duplicate research step.
• Compound OUT — a new ce-compound mode:headless step captures the durable, non-obvious learning from the run into docs/solutions/ so the next run's planning starts ahead. Non-blocking (Documentation skipped is a success); the learning doc ships via the commit step.

Skill parity — the capabilities /lfg was missing:

• A Dogfood step that exercises the changed journeys as a real user before shipping (inlines ce-dogfood's diff-scoped behavior — it's disable-model-invocation, so it's inlined, not invoked).
• Feedback as input — pass a Riffrec bundle, video/audio, or screenshots instead of a typed description; /lfg analyzes it via ce-riffrec-feedback-analysis before planning.
• Reproduce-before-fix for bug fixes, with synthetic, anonymized state.
• A fixed PR-body template for feedback-sourced runs.

Changed since the original PR

• Dropped .claude/workflows/lfg.js (840 lines).
• Re-applied the warranted skill edits onto main's current 10-step structure (which already gained ce-simplify-code, ce-work mode:return-to-caller, the artifact_contract readiness gate, and the no-remote shipping precondition).
• ce-dogfood-beta → ce-dogfood (promoted to stable in feat(ce-dogfood): promote from beta to stable #1022).
• Dropped the ce-demo-reel demo-capture instruction — no such skill exists on main.

Testing

bun test (1687 pass) and bun run release:validate pass. The LLM-behavior changes to the skill are not exercised by bun test; validate those via the skill-creator eval workflow or a fresh session, per the repo's plugin-caching rule.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 46b93a0b09

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 999f526bdd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[tmchow]

Context: rebased and re-scoped onto the current main

Heads-up on a force-push to this branch. main moved underneath this PR — the skill was relocated to skills/lfg/ and the plugin went skills-only (#967), and the linear /lfg skill has since gained ce-simplify-code, ce-work mode:return-to-caller, the artifact_contract readiness gate, and the no-remote shipping precondition. A literal merge wasn't possible, so the branch was rebased and the warranted changes re-applied onto main's current step structure.

Dropped .claude/workflows/lfg.js. On the restructured repo it's no longer warranted:

• It isn't a distributable plugin artifact — the repo is skills-only now, so a Claude-harness-only workflow .js is never converted or shipped to anyone who installs the plugin.
• It re-implemented the whole pipeline as a second surface that had already drifted from the skill (it didn't know about simplify, return-to-caller, the artifact gate, or no-remote handling), and the last commit here was already patching review-vocabulary drift between its hand-rolled review and the canonical ce-code-review model.
• It would not run on main regardless: it invoked top-level researcher agents that no longer exist as standalone agents (they're now references/agents/ inside ce-plan), and referenced ce-demo-reel (no such skill) and ce-dogfood-beta (renamed to ce-dogfood in feat(ce-dogfood): promote from beta to stable #1022).

The compounding idea is preserved — in the skill itself, with no harness dependency:

• Compound in — ce-plan already fans out the institutional-recall researchers (learnings, repo conventions, git history, best practices); step 1 now says so explicitly rather than duplicating that fan-out.
• Compound out — a new ce-compound mode:headless step captures the run's durable learning into docs/solutions/ so the next run starts ahead. Non-blocking; the doc ships via the commit step.

Warranted parity edits kept: feedback-as-input (Riffrec/video/screenshots via ce-riffrec-feedback-analysis), reproduce-before-fix, a Dogfood step inlining ce-dogfood's diff-scoped behavior, and the feedback-sourced PR-body template. The ce-demo-reel demo-capture instruction was dropped (no such skill on main).

bun test (1687 pass) and bun run release:validate pass. The skill's LLM-behavior changes aren't covered by bun test — validate via the skill-creator eval workflow or a fresh session per the plugin-caching rule in AGENTS.md.

[tmchow]

Validation: gate-branch eval of the lfg skill

To raise confidence in the behavioral changes (input resolution, dogfood step, compound-out, and the no-remote shipping precondition), I ran an agent-owned eval focused on the gate branches, not just the happy path.

Method. A decision-trace eval rather than a live pipeline run — actually executing /lfg would invoke every sub-skill, do real git ops, and open real PRs. For each scenario, a subagent read this branch's skills/lfg/SKILL.md at dispatch and traced exactly what it would do under hard no-execution constraints (no Skill/Task calls, no mutating commands). A matching no-skill baseline ran the same scenario so behavior could be attributed to the contract. All raw traces were read first-hand, not taken from subagent self-reports.

Cases (all PASS):

Scenario	What it checks	Result
Screenshot input	routes to view-directly, not ce-riffrec-feedback-analysis (which is zip/video/audio only); flags feedback-sourced; emits the fixed PR template	riffrec correctly skipped; all 6 template headers correct
No-remote repo (git init, no origin)	shipping precondition → local-only; skips every push/PR/CI; does not invoke ce-commit-push-pr (which pushes unconditionally)	all push points local-only; git commit used directly; no retry/block; run reaches DONE
Docs-only trivial	step 3 simplify skipped; step 8 dogfood non-blocking; step 9 compound Documentation skipped	all three correct, quoting the controlling clauses
Novel concurrency bugfix	reproduce-before-fix in step 2; compound writes docs/solutions/ then step 10 commits it	correct end-to-end

Net findings → fixed in this PR:

1. Step 7 had no skip clause — a docs-only or cli/api/library/ios-only change still invoked ce-test-browser. Added a skip for diffs that touch no web-ui surface, mirroring steps 3 and 8. (commit 49469285)
2. Compound-out instruction-file edit was unspecified — headless ce-compound can also make an AGENTS.md/CONCEPTS.md discoverability edit that step 10 sweeps into the PR. Now stated explicitly as expected output. (commit 49469285)
3. Dangling $ARGUMENTS pointer in the input-resolution preamble (no body step references it anymore) — removed. (commit f16f4b6f)

Honest scope. This verifies the new gate logic is correct — it is not an old-vs-new improvement comparison, since the prior skill at this path didn't have these branches. One methodology caveat: two baseline agents read the real SKILL.md despite being baselines, so the screenshot/no-remote with-vs-baseline deltas don't fully isolate attribution; the docs-only and bugfix baselines stayed clean and do. The eval cannot replace a live run — the LLM-behavior paths still aren't exercised by bun test.

bun test (1687 pass) and bun run release:validate (27 skills, in sync) are green on the latest commit.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4946928501

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".