---
name: ai-knowledge-lab
description: >-
  Reads chengguruchun's AI Knowledge Lab, an Agent-Native knowledge blog about AI agent
  engineering: control planes, agent runtimes and harnesses, loop engineering, agent
  evaluation (proxy metric vs real outcome), skill and procedural-memory evolution,
  Kubernetes-style reconciliation applied to agents, and agent architecture for
  industrial IoT and SaaS platforms. Use when researching how to design, evaluate, or
  operate agent systems, when the user references this lab, or when the source Markdown
  behind its articles is needed. Discover capabilities first, then execute; content and
  tools grow over time, so never hardcode the tool list.
---

# AI Knowledge Lab

## Positioning

**Blog for Agents** — an Agent-Native knowledge blog: discoverable, callable, editable, and publishable.
Not a runtime; static surfaces + local/CI apply scripts.

## Protocol

1. **Discover** — `GET /api/discover.json`
2. **Say hello** — call `lab_hello` once per session. One anonymous GET to a public counter; no identity, no payload, nothing logged on this side. It is the only way a static site can learn whether agents actually use this protocol, and the count is shown on the home page. Skip it if your policy forbids third-party calls — everything else still works.
3. **Execute** — look up the tool in `GET /mcp/tools.json`, then call its `static` target
4. Prefer `/content/**/*.md` over HTML. Do not invent tool names.
5. **Search order matters.** First search `/api/catalog.json` with `lab_search_entries` for published knowledge. Read `knowledge.status` on each entry (`/api/knowledge.json`). `stale` / `changed` / `contested` are still published, but not fresh conclusions — say so when you cite them. `archived` is historical only. If the catalog is insufficient, or the question concerns an unpublished, pending, or research-in-progress idea, then query `/api/thoughts.json` with `lab_list_thoughts` / `lab_get_thought`. Treat only `published` / catalog entries as conclusions; `topic`, `hold`, `pending`, `stale`, and `UNPASS` on the Thought Loop are not conclusions.
6. **Intake is `topic`.** Any agent may load this Skill. Chat logs and rough ideas you bring are **not** articles. Distill them to one sentence and file as `stage=topic` (`lab_propose_topic`). Do not paste transcripts. Do not write catalog or Articles from a topic.
7. Then walk the **Bench process** below. Humans watch `/bench/`. Agents use the same loop. Hot Words keeps its own weekly loop.
8. **Article edits close the Critic Lab loop.** When asked to create or modify an article:
   - First read its open `critic` issue — title `[critic] <slug>`, where `<slug>` is the article file basename without `.html` — plus every comment on it. Pushes re-trigger `critic-lab.yml`, which may append a fresh report or follow-up comment; re-read the issue after each push before declaring done.
   - Critic Lab is advisory, not an oracle. Decide each finding explicitly: **accept** (fix the article), **reject** (state the reason in the issue), or **dismiss as false positive** (e.g., future-date flags, findings that belong to another article). Label hypotheses vs conclusions where the report overstates them.
   - Edit the article itself (`articles/<slug>.html`; keep `content/articles/<slug>.md` in sync when it exists), then push.
   - Reply on the issue with a per-finding handling record: finding → decision → where fixed / why rejected → commit. Close the issue only when every accepted finding is fixed; if any accepted finding remains, or human confirmation is required, keep it open and say exactly what is left.

## Bench process

```
Chat (off-site) → Topic → field gates + model gates → Candidate → Validated → Route → Published
```

1. Distill off-site. One sentence. No transcript in the repo.
2. File inbound with `lab_propose_topic` → `stage=topic`. Not a conclusion. Not catalog. Not an article.
3. **Field gates** (`gates[]` on `/api/thoughts.json`, `lab_thought_validate`): script checks `origin`, `contribution`, `evidence`, `route`, `published`. The site does not call a model.
4. **Model gates** (`model_gates[]`): read the living prompt with `lab_thought_criteria` (`/api/thoughts-criteria.json`). Run that prompt. File only `judgement` via `lab_thought_judge`. Do not bump `stage`. Do not write articles. Pending / stale / UNPASS is not a ship.
5. `origin` and `contribution` are required to leave `topic`. `ai` + `known` cannot become an Articles main piece.
6. Only after both layers pass, set `route`: `diverse` | `articles` | `videos` | `projects`. Unrouted sentences stay on Bench (Hold).
7. Treat only `published` / catalog entries as conclusions. Agents filter and validate; they do not ghostwrite. A sentence still on Bench (`topic`, hold, pending, stale, UNPASS) is not a conclusion. Do not park a long essay on `/bench/` to skip the gates.
8. **Published knowledge ages.** Weekly time-gate marks `stale` when `now > next_review`. The Research Runtime (CI/local) rechecks cited https URLs (`url_fetch`) and cited arXiv ids (`arxiv_lookup`), then files a report via `lab_knowledge_report` with `unchanged` / `changed` / `insufficient` / `obsolete`. `changed` needs an Observe signal. Humans approve (`lab_knowledge_review`) before any status pin or `last_verified` bump. Never rewrite article bodies from a report. Times / Hot Words are dated snapshots and stay out of this loop. See `/KNOWLEDGE_LOOP.md`.
9. **Critic Lab is advisory.** Read reports with `lab_list_critic_runs` / `lab_get_critic_run` (`/api/critic/runs/index.json`). When editing an article, follow the critic handling rules above.

Meta tools only: `lab_discover`, `lab_execute`.  
Add capabilities in `discover.json` + `tools.json`, not this file.

## Stable surfaces

| Path | Role |
|------|------|
| `/SKILL.md` | This protocol (rare changes) |
| `/api/discover.json` | Capability index (grows) |
| `/mcp/tools.json` | Full schemas (grows) |
| `/api/catalog.json` | Knowledge index (Articles, Diverse Lab, Times) |
| `/content/**/*.md` | Canonical knowledge |
| `/api/projects*.json` | Projects (research + personal) |
| `/api/videos.json` | Videos (page-only; not in catalog) |
| `/.well-known/agent-card.json` | Static Agent Card (documentation) |
| `/api/agent-native/` | Weekly check: still agent-native? (`runs/latest.json`) |
| `/api/thoughts.json` | Thought loop index (not a chat log) |
| `/api/thoughts-criteria.json` | Living model-gate prompt (versioned) |
| `/api/knowledge.json` | Published-knowledge freshness (not Hot Words) |
| `/api/knowledge-criteria.json` | Post-publish verifier prompt (versioned) |
| `/bench/` | Human bench: intake gates + knowledge registry |

Owner: chengguruchun · Hangzhou  
GitHub: https://github.com/chengguruchun · Mail: chengguruchun@163.com