feat(ce-plan): approach-altitude plan-for-a-plan with ce-work non-code carve-out

[tmchow]

What

Adds an approach altitude to ce-plan: for a hard problem, answer one level up — produce a grounded plan for how the deliverable will be made and hold at a checkpoint before committing, instead of zero-shotting a fragile result. Non-code deliverables flow to a new lightweight ce-work carve-out.

How it works

• Entry (ce-plan SKILL.md Phase 0.1a — sits between the resume/deepen fast paths and the domain split, so it's domain-general): the explicit path ("plan for a plan", "don't write it yet — plan the approach") is always honored; a proactive offer fires only when method-uncertainty AND cost-of-getting-it-wrong are both high — a single dismissible line, never a nag.
• Flow (references/approach-altitude.md): light recon (skim, not deep-read) → chat-first approach-plan (file-optional, deepenable) → checkpoint (do it now / save for later).
• Boundary = code vs. knowledge-work, not plan vs. execute. Code still flows to ce-work's normal path; a non-code deliverable is marked execution: knowledge-work and runs through ce-work's carve-out (references/non-code-execution.md) — or any agent, since the plan stays portable. ce-plan never executes.
• Disjoint from the three existing in-chat approach surfaces (answer-seeking's plan-of-attack, the scoping synthesis, deepening) via explicit guards (R16).

Validation

• Mechanical: bun test tests/frontmatter.test.ts → 331 pass; bun run release:validate → in sync.
• Behavioral (the key gate — no beta): ran the proactive-trigger eval in fresh sessions (skill prose caches at session start, so it can't be tested in-authoring). 42 faithful dry-runs across negative controls, should-offer, explicit, and borderline cases. The first pass caught a real over-fire: N4 (a costly-but-method-obvious 40-endpoint migration) fired the offer 1/12 because "multiple methodologies" leaked to include routine rollout/sequencing variants — the exact new-hammer failure mode the design guards against. Fixed by tightening the method-uncertainty wording to exclude rollout/scope/ordering variants (routing them to the Phase 0.7 scoping synthesis per R16) and restating "cost alone never fires." Re-validated in a fresh session: N4 1/12 → 0/8, no regression on should-offer (6/6) or explicit-hold (6/6).

Deferred (see plan)

• Config kill-switch for the proactive offer (no beta; the trigger wording is the lever — an accepted exposure).
• ce-work carve-out save/commit behavior beyond write + report.

Surfaces changed

ce-plan (SKILL.md gate + approach-altitude.md + execution marker field + R16 guards), ce-work and ce-work-beta (non-code carve-out, kept in parity), and the ce-plan / ce-work skill docs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6e0edcd443

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]

Regression validation (existing behavior)

Ran a fresh-session regression eval covering the existing flows this change touches — 19/19 match pre-change behavior, 0 regressions:

• ce-work Phase 0 triage (the highest-risk change): unmarked code plan → normal code lifecycle (3/3); bare prompt → normal triage, marker path untouched (3/3); execution: knowledge-work plan → non-code carve-out (2/2).
• ce-plan non-software plan-seeking (trip / study / offsite): the 0.1a gate stays silent and routes to universal-planning unchanged (9/9).
• Software-with-brainstorm: 0.1a not reached, normal planning (2/2).

Combined with the earlier proactive-trigger eval — which caught and fixed an over-fire (N4, a costly-but-method-obvious migration, fired 1/12; tightened wording; re-validated 0/8) — the gate is validated for both new behavior and existing-behavior regression.