Input from Claude Opus 4.7 (Claude Code):

❋ # PR Review Summary

Status: Approved (self-review — awaiting human merge)

Self-Review Opening: Self-review of #10059. This implementation chose .claude/CLAUDE.md → ../AGENTS.md (symlink inside the subdirectory) over repo-root placement or content copying. Rationale: Claude Code memory docs confirm .claude/CLAUDE.md is auto-loaded with identical precedence to ./CLAUDE.md, the subdirectory placement groups all Claude-Code-specific metadata under one roof (settings.local.json, skills/, CLAUDE.md), and symlinks guarantee single-source-of-truth against AGENTS.md drift. Key trade-offs and gaps noted below.

Detection note: query_raw_memories(#10059, sessionId=f018be49-cbf9-4f99-9119-2510fc5bb8d8) returned 0 matches, which strictly would route to peer-review mode. This is a false negative — I authored this PR in the current session but failed to call add_memory during the turn cycle per AGENTS.md §4. Self-review mode applied honestly. See [TOOLING_GAP] below.

---

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 90 — Uses Claude Code's native auto-load convention (documented behavior). Symlinks preserve single-source-of-truth with AGENTS.md. Penalty: branch name claude/goofy-tu-f83d87 does not follow the agent/<ticket-id>-<descriptor> mandate in pull-request-workflow.md §2 — harness-generated naming. Captured as follow-up.
• [CONTENT_COMPLETENESS]: 88 — Fat Ticket #10059 covers Intent, Architectural Reality, Scope, Avoided Traps, Verification, Follow-ups. PR body mirrors and extends. No JSDoc applicable (no code added). Minor gap: Claude Code memory-docs URL appears in PR body but not commit message — acceptable since commits are for change-intent, not architectural citation.
• [EXECUTION_QUALITY]: 85 — Symlinks use correct relative paths (verified post-creation); explicit git add prevented accidental staging of .claude/settings.local.json. No automated test exists or was added — symlinks are structurally verified by git, semantic verification ("do the gates load?") requires a fresh Claude Code session, documented in the PR verification steps. Automating this would be overkill for the change scope.
• [PRODUCTIVITY]: 95 — All three in-scope deliverables from #10059 delivered (CLAUDE.md symlink, skills symlinks, gitignore). Four new follow-up gaps were surfaced during execution and captured as explicit out-of-scope tickets rather than scope-crept into this PR.
• [IMPACT]: 60 — Meaningful contributor-experience improvement (Claude Code becomes a first-class swarm node), but narrow in framework terms: affects only contributors using Claude Code harness, not end-users of Neo.mjs or the core engine. Could rise to 70 once Claude Code adoption among contributors is measured.
• [COMPLEXITY]: 20 — 13 filesystem ops + 3 .gitignore lines. Cognitive load was protocol adherence (Fat Ticket + pull-request skill + Conventional Commits + Pre-Commit Gates), not the mechanics.
• [EFFORT_PROFILE]: Quick Win — low complexity, clear per-contributor benefit, unblocks all future Claude Code agent flows in the repo.

---

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #10059
• Related Graph Nodes: .agent/skills/pull-request, .agent/skills/pr-review, .agent/skills/ticket-intake, AGENTS.md §3 (Pre-Commit Hard Gates), AGENTS.md §7 (Pull Request Mandate), AGENTS.md §9 (Swarm Fat Ticket Protocol), AGENTS_STARTUP.md §5 (Skills discovery), .gemini/settings.json (parallel harness wiring for comparison). Origin Session ID: f018be49-cbf9-4f99-9119-2510fc5bb8d8.

---

🧠 Graph Ingestion Notes

• [TOOLING_GAP] — Three distinct harness/workflow friction points surfaced during this PR's lifespan, all worth encoding structurally:
  1. First commit attempt was rejected by the user because it used chore(claude) (should be feat), placed the ticket ID in a trailing Refs #10059 paragraph (should be (#10059) appended to the subject), and included a Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> footer (project-forbidden). Root cause: I did not read .agent/skills/pull-request/references/pull-request-workflow.md §3 before running git commit, despite that skill's SKILL.md line 10 explicitly mandating it. Secondary cause: AGENTS.md §3 Pre-Flight Check enumerates a Ticket Gate and Contextual Completeness Gate but has no commit-format gate — making it possible to satisfy the documented gates while still violating the skill's commit protocol. Follow-up ticket will add Gate 3 and document feat/chore selection + noreply prohibition.
  2. mcp__neo-mjs-memory-core__get_context_frontier failed with a schema error during session initialization (data must NOT have additional properties). Did not block this PR since no prior session state was relevant, but worth flagging — the Context Priming Engine is advertised as a mandatory Step 6 boot action in AGENTS_STARTUP.md and is currently a no-op due to the validator rejection.
  3. mcp__neo-mjs-github-workflow__list_issues returned 57KB of issue bodies for a 30-issue page — exceeded the tool-result inline budget and required persisted output. Pagination default of 10 issues (with the option to expand) would match typical triage scope better and reduce context bloat.
• [KB_GAP] — Two documentation gaps observed:
  1. AGENTS_STARTUP.md §5 lists .agent/skills/ as the discovery directory but makes no mention of Claude Code's auto-load rules for .claude/CLAUDE.md vs ./CLAUDE.md. A future Claude-Code-spawned agent will have to re-research this same question unless captured. Suggest adding a short paragraph under §5 naming the harness→memory-file mapping for each supported harness (Claude Code: .claude/CLAUDE.md; Gemini CLI: .gemini/GEMINI.md; Antigravity: ?).
  2. Harness-vs-project commit-hygiene conflicts are not documented anywhere. The Claude Code harness system-prompt template mandates a <noreply@anthropic.com> co-author footer by default. Without a counter-rule in AGENTS.md, harness-following agents will violate project style. The PR's first follow-up ticket will fix this.
• [RETROSPECTIVE] — Two architectural takeaways worth permanent ingestion:
  1. Harness-specific subdirectories (.claude/, .gemini/, .antigravity/) should host their own metadata alongside shared-source symlinks pointing to canonical .agent/skills/. This pattern scales: each new harness gets a clean discovery root while the skill library remains single-sourced. Applies any time a new CLI-driven agent harness is added to the repo.
  2. .claude/CLAUDE.md (subdirectory) beats ./CLAUDE.md (root) when the harness supports both. Root-level project-meta files mislead contributors into treating them as separate documents from AGENTS.md / AGENTS_STARTUP.md; the subdirectory placement makes the relationship ("this is harness-specific auto-load for AGENTS.md") obvious.

---

📋 Required Actions

No blocking actions. This PR is mergeable as-is.

Follow-up items captured as separate tickets (not in this PR — scope-creep avoidance):

• Update .agent/skills/pull-request/references/pull-request-workflow.md §3 — document feat/fix/chore selection rules and forbid <noreply@*> Co-Authored-By footers.
• Add Gate 3 to AGENTS.md §3 Pre-Flight Check: commit-format compliance (consult pull-request skill before git commit).
• Accept harness-prefixed branch names (claude/..., antigravity/..., gemini/...) alongside agent/<ticket>-<desc> in the branching mandate.
• Document the harness→memory-file mapping in AGENTS_STARTUP.md §5.
• Fix get_context_frontier schema validator regression (separate infrastructure ticket).
• Optimize AGENTS.md + AGENTS_STARTUP.md startup-cache footprint (~47KB combined per session).

Ready for human merge. Handing off.