Migrate skill routes "AI collaboration rules" to harness memory/feedback_*.md, contradicting v5 self-healing rule

[NodarDavituri]

The Migrate skill (skills/Migrate/SKILL.md) routes chunks classified as "AI collaboration rules" to ~/.claude/projects/{harness-dir}/memory/feedback_*.md. The v5 system prompt's "Self-Healing Infrastructure" section explicitly says that surface is the wrong home for rules/preferences/operational behavior:

"For rules, preferences, and operational behavior, ignore [harness auto-memory] guidance ... Every 'feedback memo' is a missed system patch."

Issue #1054 was closed on 2026-05-01 with the comment that v5 documentation surfaces this distinction explicitly. However, the Migrate skill was last edited 2026-05-02 and still writes to memory/feedback_*.md. Three locations encode the conflict:

• Frontmatter description: "AI rules (memory/feedback_*.md — new file per chunk)"
• Destinations table (SKILL.md line 38): AI collaboration rules → memory/feedback_*.md (for "always do X", "never Y" patterns)
• Rules section (SKILL.md line 140): "Feedback memories get new files. Each memory/feedback chunk becomes its own feedback_migrated_<slug>_<id>.md file — not appended to an existing memory."

Reproduction

1. Run /migrate against any source containing AI-rule-shaped chunks (Cursor rules, old feedback_*.md files from a prior install, CLAUDE.md operational rules sections).
2. The classifier labels them as feedback.
3. MigrateApprove.ts writes new feedback_migrated_<slug>_<id>.md files into harness memory.
4. Result: rules land at the exact path the v5 constitutional rule designates as "the wrong home."

Suggested resolution

Migrate's "AI collaboration rules" classification should not auto-commit to harness memory. Each such chunk should be flagged for human routing across the four constitutional surfaces — CLAUDE.md Operational Rules, a hook, settings.json, or a skill SKILL.md Gotchas section — depending on the rule's nature.

Concretely:

• Keep the feedback classification in MigrateScan.ts so chunks are still detected.
• In MigrateApprove.ts, replace the auto-commit branch for feedback-classified chunks with a walk-through prompt that asks the user which constitutional surface should host each rule, then performs an Edit/Write into the chosen target file rather than into harness memory.
• Update SKILL.md frontmatter, destinations table, and rules section to match.

Context

Found while extracting feedback files from a v4 backup into v5 — would have run MigrateApprove and re-created exactly the violation the v5 rule forbids. Caught at the read-and-classify step by reading the three v4 feedback_*.md files directly and noticing two of them were already structurally encoded in v5 skills (the v5 system had self-healed against the doctrine; the Migrate skill had not).

Environment

• PAI v5.0.0
• Algorithm v6.3.0
• skills/Migrate/SKILL.md mtime: 2026-05-02
• Constitutional source: PAI/PAI_SYSTEM_PROMPT.md § "Self-Healing Infrastructure" / "Override of harness auto-memory"

[danielmiessler]

You caught a real doctrine violation and it's still there — the Migrate skill never got updated after the self-healing rule landed, so it still routes rules to the exact surface the system prompt calls a missed system patch. We're adopting your suggested resolution: keep the classifier, replace the auto-commit with human-routed placement into the constitutional surfaces. The detail that your own v4 files had already been structurally absorbed by v5 skills was the perfect proof of the doctrine.