8e24d0694a
feat: add chainTo skill chaining, prompt signals, validation upgrades, and dedup reset on context clear
...
Add a chainTo frontmatter field to skills that triggers follow-up skill injection
when PostToolUse file contents match regex patterns. Add upgradeToSkill/upgradeWhy
fields to validation rules so validate errors can recommend loading a specific skill.
Register posttooluse-bash-chain.mjs in hooks.json. Add ChainToRule interface to
patterns.mts and skill-map-frontmatter.mts, with parseChainToRules() parser. Add
duplicate-key detection to the YAML parser. Reset dedup claim dir and session file
on clear/compact events in session-start-seen-skills so skills re-inject into fresh
context windows.
Add chainTo rules to: agent-browser-verify, agent-browser, ai-elements, ai-gateway,
ai-generation-persistence, ai-sdk, auth, bootstrap, chat-sdk, cms, cron-jobs,
deployments-cicd, email, env-vars, geist, investigation-mode, json-render,
marketplace, micro, ncc, next-forge, nextjs, observability, payments,
react-best-practices, routing-middleware, runtime-cache, satori, shadcn,
sign-in-with-vercel, swr, turbopack, turborepo, v0-dev, vercel-agent, vercel-api,
vercel-cli, vercel-firewall, vercel-flags, vercel-functions, vercel-queues,
vercel-sandbox, vercel-storage, verification, workflow. Add upgradeToSkill to
ai-elements and ai-sdk validate rules. Expand ai-sdk validate messages with
Run Skill() hints. Update nextjs, vercel-storage, runtime-cache, workflow, turborepo
skill bodies.
Add new skills: geistdocs (Geist design system docs), zzz-test-meta-name-mask
(test fixture). Add skills/_chain-audit.md chain coverage audit doc.
Delete .claude-plugin/marketplace.json, .claude-plugin/plugin.json (deprecated),
skills/edge-runtime/SKILL.md (consolidated into vercel-functions).
Add tests: posttooluse-chain.test.ts (4699 lines, chain injection e2e),
ai-sdk-companion.test.ts (181 lines). Expand build-skill-map.test.ts (+335 lines),
validate-rules.test.ts (+936 lines), session-start-seen-skills.test.ts (+74 lines),
skill-map-frontmatter.test.ts (+50 lines), verification-skill.test.ts (+20 lines).
Update build-manifest.ts to emit chainTo rules and upgradeToSkill fields. Rebuild
generated/skill-manifest.json, generated/skill-catalog.md, generated/build-from-skills.manifest.json.
Rebuild all compiled hooks/*.mjs. Update CLAUDE.md lexical prompt default to on.
Update vercel.md ecosystem graph, docs, and cli-reference.
2026-03-11 15:18:35 -06:00
7ac8edfb8f
feat(skills): aggressive prompt signals for workflow/sandbox, "check the docs" directives across 8 skills; bump to 0.11.0
...
- workflow: add ~60 new promptSignal phrases for reliability language
(survive page reload, fault-tolerant, retry on failure, session
persistence, reconnect, durable chat/agent), human-in-the-loop
patterns (approval, pause until, wait for), and pipeline vocabulary.
Add ~25 new allOf pairs and 12 new anyOf terms.
- vercel-sandbox: add promptSignals from scratch (had none). 33 phrases
covering isolated execution, sandbox environments, code safety,
FFmpeg/media processing, code playgrounds, and tutor patterns.
27 allOf pairs, 8 anyOf terms, noneOf excludes iframe/codesandbox.
- Add "CRITICAL — your training data is outdated" blockquote at the top
of 8 fast-moving skill bodies (workflow, ai-sdk, vercel-sandbox,
ai-gateway, chat-sdk, vercel-flags, vercel-queues, ai-elements)
directing agents to fetch docs before writing code.
Eval results that motivated these changes:
- content-pipeline: 9/10 (workflow triggered, correct WDK patterns)
- customer-support: 3/10 → workflow never injected (now scores 50)
- code-sandbox-tutor: 1/10 → sandbox never injected (now scores 80)
2026-03-09 13:16:16 -06:00
John Lindquist