feat: route Sume agent skills for SUM-619

[chasehuh]

Summary

• Add agent skill routing guidance to AGENTS.md, README, and local docs.
• Rewrite the sume router skill so agents can route by intent before using focused skills.
• Add focused bundled skills for auth, minimum-latency iteration, quality experiments, spend-capped dogfooding, and local media QA.
• Refine existing tools/avatar/avatar-video/assets skills with schema-first guidance, Not For boundaries, paid/idempotency/cap language, and current MCP coming-soon wording.
• Regenerate the embedded bundled-skill fallback and add regression tests for discovery/routing safety.

Linear

• SUM-619: cli / skill 개선
• SUM-624: maximize concurrency / spend cap validate

Skills Added/Updated

Added:

• sume-auth
• sume-min-latency
• sume-quality-experiment
• sume-spend-capped-dogfood
• sume-media-qa

Updated:

• sume
• sume-tools
• sume-avatar
• sume-avatar-video
• sume-assets

Validation

• pnpm vitest run test/skills-registry.test.ts test/cli.test.ts
• pnpm run check
• pnpm run build:binary
• ./dist/sume --version
• ./dist/sume --json skills list
• ./dist/sume --json skills export sume
• ./dist/sume --json tools schema avatar-videos.create
• binary skill smoke script verifying new bundled skills and router export

Note: pnpm exec prettier --write ... was attempted, but this repo does not include Prettier, so no formatting tool was run. git diff --check passed.

Known Limitations / Follow-ups

• MCP remains explicitly coming soon in this PR; no MCP command behavior is exposed or changed.
• This PR does not add a hosted skill registry or change release/deploy behavior.
• Future work can add live no-cost/paid dogfood QA only after explicit cap/authorization.

Safety Confirmation

• No paid/provider calls.
• No deploy, release, publish, or tag.
• No production/env/DB/cloud mutation.
• No secrets, API keys, signed/private URLs, or raw provider payloads added.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1e70215a4d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Add idempotency to the fast paid-submit example

For the single-job minimum-latency path, this paid avatar-videos create example omits --idempotency-key, so following it after a timeout or retry can submit a second paid job despite rule 5 above requiring stable keys. I checked the CLI contract in src/commands/avatar-videos.ts, where --idempotency-key is an explicit option, and idempotencyHeaders() only forwards options.idempotencyKey, so this command does not add a stable key automatically.

Useful? React with 👍 / 👎.