[kalwalt]

Closes #7.

Summary

AGENTS.md stays the single source of truth. Each tool gets a thin entry point that points to it, and the review checklists are shared by all tools.

Tool	Entry point
Codex, Cursor, Antigravity	read AGENTS.md natively (no file needed)
GitHub Copilot	.github/copilot-instructions.md (critical rules inlined, for chat and code review on github.com) and .github/instructions/{graph,tracking}.instructions.md (path-specific, applyTo)
Gemini CLI	.gemini/settings.json → context.fileName: ["AGENTS.md"]
Claude Code	CLAUDE.md (Claude-only notes), .claude/agents/{graph,tracking}-reviewer.md, .claude/skills/{new-node,check-upstream}

Other changes:

• AGENTS.md: which tools read it, adding a node, test assets and the fake camera, review checklists.
• docs/agents/checklists/: graph-review.md and tracking-review.md, tool-agnostic. The tracking checklist includes readiness for the jsartoolkitNFT facade (Adopt the jsartoolkitNFT data-oriented tracker facade (jsartoolkitNFT#720) #5).
• scripts/check-agent-docs.mjs (npm run check:agents, run in CI). It fails when:
  • a pointer file stops linking to AGENTS.md;
  • an agent file names an npm run script that does not exist;
  • an agent file names a path, a relative link or an applyTo glob that does not exist.
• CONTRIBUTING.md: a paragraph on keeping the agent files in sync.

Issues, milestones and releases

Added at the maintainer's request:

• AGENTS.md → "Planning and tracking work":
  • one milestone per release;
  • how to implement an issue (branch <type>/<n>-<slug> from origin/dev, PR with Closes #n);
  • when an issue is done;
  • a milestone is closed only after its release exists.
• Why issues stay open: the default branch is master and PRs target dev, so Closes #n takes effect only when the release reaches master.
• docs/agents/checklists/milestone-audit.md: checks each issue of a milestone on evidence, not on its state:
  • a merged PR into dev;
  • the task list ticked;
  • the acceptance criteria met;
  • the sub-issues done;
  • CI green and the CHANGELOG updated.
• Claude Code:
  • milestone-auditor agent (read-only);
  • work-on-issue skill;
  • release skill: audit, version and CHANGELOG, dev → master with Closes lines, tag and GitHub release, then close the milestone. Every outward step needs the maintainer's confirmation.
• Copilot instructions and CONTRIBUTING.md: the same rules, briefly.

Deviations from the issue

• No .agents/instructions.md. Antigravity reads AGENTS.md natively; its documented workspace rules location is .agents/rules/*.md, not .agents/instructions.md.
• No GEMINI.md. The current Gemini CLI setting is context.fileName.
• Added Copilot path-specific instructions and the check-upstream skill, as proposed in the comment on Agent instructions for AI coding tools (Claude Code, Copilot, Gemini, Antigravity, Codex, Cursor) #7.

Test plan

• npm run typecheck
• npm test: 44 tests pass
• npm run build
• npm run check:agents passes (16 files), and fails as expected on three breakages:
  • a pointer file without AGENTS.md;
  • npm run lint named in AGENTS.md;
  • a misspelt src/runtime/tracker/ path.
• graph-reviewer on a test diff renaming detectionInterval → interval without a GRAPH_VERSION bump: Blocking (no bump, stale public/graphs/, no CHANGELOG line).
• tracking-reviewer on a test diff moving lost-debounce logic into the worker: Blocking (worker holds state, unbalanced found / lost).
• milestone-auditor on v0.1.0 — Foundation: Not ready (PR docs: add agent instructions for AI coding tools #11 not merged, feat: tracking traces, replay and debugging tools #9 not started), as expected.
• npm run test:e2e: passed in CI (the test job); not run locally.

🤖 Generated with Claude Code