feat(lfg-beta): add beta variant with Codex delegation and provider hints

[davidalee]

Summary

Introduce lfg-beta as a parallel beta of /lfg, mirroring how ce-work-beta parallels ce-work. Stable /lfg is unchanged.

What's new in the beta:

• delegate:codex / delegate:local tokens (and lfg_beta_delegate config) route the work phase through ce-work-beta delegate:codex.
• providers:<list> token (and lfg_beta_providers config) accepts arbitrary provider names. lfg-beta auto-detects each name across three surfaces — mcp__<name>__* tools, CLIs on PATH (command -v), installed skills — and passes a providers_hint: line to ce-plan and the work skill so they can prefer those helpers. No registry of per-provider behaviors; codex and perplexity have inline wiring at specific phases, everything else flows through the generic hint path.
• Parallel codex-MCP adversarial review lane runs alongside ce-code-review in step 3 when adversarial_review_active is true and the codex MCP is loaded. Findings flow through the standard residual handoff with synthesized owner: downstream-resolver / source: codex-adversarial fields.
• Residual handoff gains an inline [needs human review] marker for not-autofixed judgment-required findings, plus an Items applied with low confidence subsection for autofixed-but-uncertain items, so anything you might want to redirect is easy to spot.

Files

• New skill: plugins/compound-engineering/skills/lfg-beta/SKILL.md + references/codex-adversarial-prompt.md + references/tracker-defer.md (the last copied from lfg/ since each skill is self-contained per AGENTS.md).
• .compound-engineering/config.local.example.yaml documents the three new lfg_beta_* keys.
• plugins/compound-engineering/README.md adds the /lfg-beta row to the Beta / Experimental table.
• tests/skill-agent-ce-prefix.test.ts exempts lfg-beta from the ce- prefix rule (matches the existing lfg exemption with the same justification).

Why beta and not in-place

Same reasoning as ce-work-beta: feature involves a new external dependency (codex CLI/MCP) and a new config surface. Beta lets users opt in via /lfg-beta while stable /lfg keeps its current orchestration.

Promotion path is the standard one: drop the -beta suffix, drop disable-model-invocation: true, drop the [BETA] description prefix, swap the README rows, remove the test exemption.

Test plan

• bun test (1250 pass, 0 fail)
• bun run release:validate (sync clean — 39 skills now)
• tests/skill-shell-safety.test.ts and tests/frontmatter.test.ts pass for the two new pre-resolution blocks
• Behavioral validation via skill-creator — per plugins/compound-engineering/AGENTS.md, skill prose changes need skill-creator evals. Recommend running before public exposure on at least: token parsing (delegate, providers, normalization edge cases), delegation routing branch, parallel adversarial dispatch.

Items needing your judgment

1. Token name providers: vs separate mcp: / cli:. The original ask used mcp:perplexity. I generalized to a unified providers: since names commonly span surfaces (codex is both CLI and MCP) and the skill auto-detects which surface(s) a name resolves to. Tradeoff: less precise than letting the user say "I specifically want the MCP variant of foo". Easy to add mcp: / cli: precision later if needed.

2. /lfg-beta cannibalizing /lfg. README places both rows side-by-side. There's no auto-routing — model invocation is disabled on both, so the user picks. If you'd like /lfg to nudge users toward /lfg-beta (or vice versa), that's a separate change.

Closing PR #13 (the in-place lfg modification approach) — superseded by this beta path.