Sweep over agent skills plus a repair of the skills-publishing pipeline, which has been silently emitting an empty registry since the skills directory moved.

Skill doc fixes

• error-pulse: removed the iterate tool subscribe-slack-thread / get-current-session-id instructions — those tools were deleted with the daemon stack in Remove legacy OS1 stack (apps/os, daemon, sandbox) #1341 and no equivalent thread-subscription mechanism exists today. The skill now states that escalation-thread follow-up is manual.
• architect:
  • Required MCP access now points at the general Cloudflare API MCP server (https://mcp.cloudflare.com/mcp) instead of the product-specific observability endpoint that .agents/skills/cloudflare-traces/SKILL.md and .agents/skills/debug-mcp-server/SKILL.md explicitly forbid.
  • Dropped the fly CLI from diagnosis tool access and de-Fly'd the monitoring-mode section (kept as a pointer to architect-monitoring, which now covers OS/Cloudflare health monitoring; the Fly stack is removed).
  • Replaced the dead iterate tool exec-ts 'await slack.chat.postMessage(...)' snippet with a plain instruction to post the summary to #monitoring, noting the daemon tooling is gone. There is no documented programmatic Slack-posting CLI in the repo today, so the instruction stays mechanism-agnostic.
• architect-diagnose-performance: removed the "Fly signals (machine health/usage)" input.
• playwriter-spec: fixed the run command — apps/os has an e2e script (no test:e2e) and the suite reads APP_CONFIG_BASE_URL (not OS_BASE_URL).
• cloudflare-traces EXAMPLES.md: the os-preview-2 example now uses the dev/preview account id 376ef7ed81b0573f93524de763666c15 instead of the garple account.
• debug-os-worker: apps/os/src/entry.workerd.ts no longer exists; replaced with apps/os/src/worker.ts.

Registry pipeline repair

• apps/iterate-com/scripts/generate-skills-registry.mjs still read <repoRoot>/skills, which was moved to .opencode/skills in fix: Move skills to the correct place #1097, and swallowed the ENOENT — so the published .well-known/skills registry has been empty ever since. The generator now reads .opencode/skills and fails loudly if the directory is missing.
• Regenerated apps/iterate-com/backend/generated/skills-registry.ts; it now contains the two publish: true skills (example-skill, playwriter-spec).
• lint-staged.config.cjs had the same stale skills/** glob, so the auto-regeneration hook never fired on skill changes; updated to .opencode/skills/**.
• .opencode/skills/AGENTS.md updated to describe the fixed pipeline.
• example-skill kept publish: true and got an honest description/body as a minimal registry smoke-test skill (see flags below).

Checks

• pnpm install, pnpm format (oxfmt), pnpm lint, pnpm typecheck — all green.
• node apps/iterate-com/scripts/generate-skills-registry.mjs and pnpm --dir apps/iterate-com skills:generate — output contains both publish:true skills.
• pnpm --dir apps/iterate-com build — builds successfully with the regenerated registry.

Skipped

• Nothing skipped; all nine items re-verified against the code and applied. Did not touch .agents/skills/adding-a-new-durable-object-mixin/ or .agents/skills/drizzle-migrations/ (owned by other PRs).

Flags for reviewers

• example-skill judgment call: I kept publish: true and rewrote it as an explicit registry smoke-test skill rather than unpublishing it — having one always-published skill gives the .well-known/skills endpoint a guaranteed non-empty payload to check against. Flip to publish: false if you'd rather not ship it.
• Architect Slack reporting is now mechanism-agnostic: with the daemon tooling gone I found no replacement programmatic Slack-posting path in the repo, so the skill just says to post to #monitoring with whatever access the runtime has. If a real mechanism exists (or gets added), the skill should name it.
• Merging this changes what iterate.com/.well-known/skills serves on next iterate-com deploy: from an empty list to example-skill + playwriter-spec.

🤖 Generated with Claude Code