prompt(skill+memory): enforce sequential destructive pair + final-text honesty

[RowitZou]

Summary

Closes the prompt-side root causes (#2 + #6) from the 2026-05-26 dogfood Bug 1 — companion to #11 which physically blocked the silent data loss with a same-name destructive guard.

The dogfood signal: skillConsolidator emitted SkillWrite(name:X, overwrite:true) + SkillDelete(name:X) in one batch, SkillWrite failed validation, SkillDelete ran anyway, user's prior on-disk skill silently gone. Model final text: "nothing to merge this pass" — it didn't even notice it had destroyed user data.

Prompt changes

skillConsolidator.ts

Step 2 — explicit "single skill is never a merge candidate" guard. Today's degenerate case was a size-1 group treated as X merges into X.

Step 3 split into 3.a / 3.b / 3.c with hard sequencing:

• 3.a SkillWrite is a standalone tool_use; wait for is_error:false
• 3.b deletes only run after that confirmation, and never delete the surviving name
• 3.c on is_error:true, abort the group; no blind retry
• "Do not bundle 3.a and 3.b in the same assistant turn"

Output discipline — final text must enumerate SkillWrite + SkillDelete success counts plus any aborted groups. Explicit Never say "nothing to merge" if you called any SkillWrite or SkillDelete — that exact phrase is what the model used while it had just deleted the user's only skill.

memoryCurator.ts

Same sequencing applied to the MemoryWriteAt + MemoryDelete pair:

• Cross-role promotion: "first MemoryWriteAt ... as a standalone tool_use, wait for is_error:false, then MemoryDelete the originals. If the MemoryWriteAt returns is_error:true, abort the promotion — do not delete any originals."
• Within-directory cleanup: same chain
• Output discipline: enumerate destructive op counts; never claim "nothing changed" when destructive tools fired

Test plan

• Snapshot test hashes (role-prompt.snapshot.test.ts + prompt/prompt-for-role.test.ts) updated for the 2 roles — deliberate semantic change, not a structural refactor (per [feedback_prompt_changes]).
• New curator-prompt-contract.test.ts (13 cases) anchors the key sequencing phrases as intent guards on top of the snapshot's byte pin. If a future rewrite forgets the sequence-after-success rule, contract tests fail loudly with which phrase is missing.
• pnpm test 1724/1724 PASS (1711 baseline + 13 new contract cases)
• pnpm typecheck clean
• pnpm build clean
• Manual: dogfood — let the curators run a dream pass; confirm skillConsolidator no longer emits same-name SkillDelete, and confirm final text now enumerates op counts honestly

Relationship to #11

	#11 (merged-ready)	This PR
Layer	tool / parser / observability	prompt
Mechanism	(userId, name) TTL refuses same-name delete after recent write failure; flow-style YAML now parses; actionable error msgs; stderr	Tells the model to sequence the pair and own up to destructive ops
Why both	Defense-in-depth: prompt makes the right thing easy, guard catches the wrong thing

After both ship, the dogfood failure cannot recur even if any one layer regresses.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates the bundled skillConsolidator and memoryCurator role prompts to prevent destructive “write+delete” pairs from executing out of order (or after a failed write), and tightens final-text reporting so the agent acknowledges destructive operations.

Changes:

• Update skillConsolidator prompt to require sequential SkillWrite → (only on success) SkillDelete, forbid same-name delete, and improve “nothing to merge” honesty rules.
• Update memoryCurator prompt to require sequential MemoryWriteAt → (only on success) MemoryDelete, and add honest final-text reporting rules.
• Update snapshot hashes and add a new contract test suite to pin key sequencing/intent phrases beyond byte-level snapshots.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/agents/bundled/skillConsolidator.ts	Adds explicit sequencing + safety rules for SkillWrite/SkillDelete and more honest final output requirements.
src/agents/bundled/memoryCurator.ts	Adds explicit sequencing rules for MemoryWriteAt/MemoryDelete and honest final output requirements.
src/agents/bundled/curator-prompt-contract.test.ts	New intent/contract tests to ensure critical sequencing phrases remain present in prompts.
src/agents/role-prompt.snapshot.test.ts	Updates expected snapshot hash for prompt semantic changes.
src/prompt/prompt-for-role.test.ts	Updates expected snapshot hash for prompt semantic changes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.