feat(ai): add pipeline driver skill (plan/implement/review) - #16

Problem

Multi-stage work (investigate/plan, implement, review-fix loop) has only been orchestrated by prompts ($pr-closeout, lfg-style skills), which do not hold gates and are tied to one harness. The goal is a deterministic driver that works the same whether the main agent is Claude Code, Codex, or Amp, runs each stage on a chosen headless engine, and returns control to the main agent between stages.

Changes

New pipeline skill (modules/ai/skills/pipeline, profile coding):

• scripts/pipeline: stdlib-only driver with new, plan, implement, review, run [--until], status, summary, reject. Each stage is one fresh engine process (codex, claude, or grok) with a JSON schema; results steer control flow. Exit codes: 0 done, 1 error, 3 needs the main agent.
• new validates flags and the autoreview helper before creating an isolated jj workspace or git worktree (or checks that --workspace belongs to the repo). Runs live under ~/.local/state/pipeline/<run> with task.md, plan.md/json, implement.json, review-N.json, fix-N.json, summary.md, prompts, and full engine logs.
• Plan stage is read-only per engine (codex read-only sandbox; claude plan mode with read tools, no user MCP servers; grok --tools allowlist plus MCP meta-tool deny). Every write stage refuses to start on a dirty workspace. Implement/fix run with bypass inside the isolated checkout, as in $codex-first. Codex runs --ephemeral; grok sessions are preassigned and deleted afterwards.
• Review stage wraps autoreview --reviewers ... (default codex,grok) on the frozen branch, then a fix round (fix(review): commits), re-review. --max-rounds (5) caps fix rounds, not reviews, so a rerun after the main agent's own fixes can still confirm clean. Implementer rejections carry their reason into the next review prompt; a finding surviving two fix rounds halts; pipeline reject records the main agent's own rejection. Round and survival counters persist only after a fix completes.
• Halts (exit 3): plan open_questions; implement/fix question|blocked|scope_change; failing tests; done without a new commit; dirty workspace after a stage; incorrect verdict without findings; round cap. run resumes from the last completed stage and retries the halted one.
• Defaults: plan claude:fable:high (fable = latest Fable alias, 5.1 today) in plan mode, implement codex:gpt-5.6-sol:high, review codex:gpt-5.6-sol:xhigh,grok:grok-4.6:xhigh; inside an Amp orb the review default is the same panel through amp (amp:openai/gpt-5.6-sol:xhigh,amp:xai/grok-4.6:xhigh).
• summary templates summary.md from the artifacts with no LLM call.
• Prompts in prompts/*.md; SKILL.md documents the contract; skill-overrides.json publishes it to the coding profile.

Tests

• End-to-end on a toy git repo: plan (grok and claude engines) -> implement (codex) -> review panel clean; produced commits pass uvx pytest -q and the CLI works.
• Fix loop with a stub reviewer: finding -> fix(review) commit -> clean; already-addressed finding -> rejected without commit -> clean.
• Resume after a disk-full engine crash: run skipped the completed plan and retried implement.
• Unit checks via import: single-pass template rendering, failed-test/no-commit/status halts, --until on completed stages, review-before-implement rejection, invalid engine leaves no workspace, workspace ownership for jj workspaces and git worktrees, sibling helper resolution through an installed symlink.
• make test (darwin build) passes with the new skill directory.
• autoreview --reviewers codex,grok on this change ran the full 5-round cap. Rounds 1-4 findings were all fixed. Round 5 reported 6 findings; 5 were fixed after the cap without a further review round (claude planner moved to plan mode without Bash allow-globs and with --strict-mcp-config; pre-stage dirty check for write stages; reviewer spec validated at new; fix-round and survival counters persisted only after a completed fix), and 1 was rejected: "resume a halted fix before accepting a later clean review" is the main agent's job after a failing-test halt, now stated in SKILL.md. The last review round therefore has not been re-run; see the PR discussion if you want another pass.