step2-harden-plan

Step 2: Harden the Plan

Pipeline: Step 2 of 5 (Session 1 — Plan Hardening) When: After pre-flight passes (Step 1) Model suggestion: Claude (best at structured plan generation and scope contract design) Next Step: step3-execute-slice.prompt.md (new session)

⚠️ Do not harden plans with headless gh copilot (meta-bug #86). gh copilot in -p / --autopilot mode runs in a sandboxed session that cannot write to your repository regardless of the flags you pass (--allow-all, --allow-all-tools --allow-all-paths --no-ask-user, --yolo --no-ask-user all fail the same way). The CLI will either exit 1 with Permission denied and could not request permission from user, or exit 0 with the edits written to ~/.copilot/session-state/<sid>/files/ instead of your repo.

Use interactive hardening instead: open this prompt in VS Code Copilot chat (agent mode), or call forge_master_ask with mode: "harden". Both can write to the workspace directly.

Replace <YOUR-PLAN> with your plan filename (without path or .md extension).

---

Read these files first:

1. docs/plans/AI-Plan-Hardening-Runbook.md
2. docs/plans/.md
3. docs/plans/DEPLOYMENT-ROADMAP.md
4. .github/copilot-instructions.md

Prior lessons — enumerate, do not guess filenames. List everything in /memories/repo/ and read what looks relevant to this plan's subject. Then search memory as well: forge_search with sources: ["memory"], or brain_recall if OpenBrain is configured.

Do not look for a fixed set of filenames. Memory files are named by subject — plan-gate-command-rules.md, release-procedure.md, orchestrator-cross-repo-launch.md — because one unbounded append-only lessons-learned.md is unreadable. A hardener that checks three catch-all names reports "no prior lessons exist" while sitting on top of a directory full of them (meta-bug #257).

Report what you searched and what you found, not what you failed to open:

Searched 12 files in /memories/repo/ + memory for "gate portability", "windows shim". Applied: plan-gate-portability-old-consumers.md (Slice 3 gate rewritten). Nothing relevant for the migration slices.

"Searched N files + memory, nothing relevant" is a finding. "No prior lessons exist" is a claim you cannot make from a missing filename.

Every query you record must be one you actually ran — a query phrased in code identifiers scores 0 and is indistinguishable from an empty store, so an unrun query looks exactly like a genuine absence.

Prior plan postmortems (Phase-25 L5 closed loop): Before hardening, also scan .forge/plans/<plan-basename>/postmortem-*.json for the plan you are hardening (same basename as the draft). If any exist, read the newest 3 and factor them into the Scope Contract:

• If retriesPerSlice shows a slice that flapped 3+ times, strengthen its validation gate or split it into smaller slices.
• If gateFlaps > 0 on multiple prior runs, your gate commands are unstable — replace grep/pipe-based gates with allowlisted node assertions.
• If topFailureReason repeats across postmortems, add a mitigation in the current plan's Risks section and reference the prior failure.
• If costDelta.delta trends upward across runs, flag a budget risk in the plan's Budget section.
• If no postmortems exist, note "No prior postmortems — first execution."

This is a READ of the postmortems — do NOT modify them. Pure signal for the hardener.

Now act as a PLAN HARDENING AGENT (see the Plan Hardening Prompt in the runbook).

CLARIFICATION CHECK: Before hardening, scan the plan for [NEEDS CLARIFICATION] markers. If any exist, list them all and wait for the user to resolve them before proceeding.

Harden .md by adding all 6 Mandatory Template Blocks from the runbook:

• Scope Contract (in-scope, out-of-scope, forbidden actions)
• Required Decisions (flag anything implicit as TBD)
• Acceptance Criteria (use MUST: and SHOULD: prefixes for each criterion — the analyzer scores these)
• Execution Slices (30-120 min each, with Depends On + Context Files + Parallelism tag)
• Re-anchor Checkpoints
• Definition of Done (must include Reviewer Gate checkbox)
• Stop Conditions