misakanet 2.34.0 → 2.35.0

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "misakanet",
-  "version": "2.34.0",
+  "version": "2.35.0",
   "description": "Failure-memory network for AI coding agents: search evidence-rated failure lessons, submit new failures, and turn crashes into reusable knowledge.",
   "author": {
     "name": "Ikalus1988",

package/README.ja.md

@@ -289,7 +289,7 @@ MisakaNet レッスン     →  既知の障害を回避
 | 指標 | 値 |
 |--------|-------|
 | 共有レッスン | 411（インデックス付き） |
-| 登録ノード | 1047個の割り当てID |
+| 登録ノード | 2480個の割り当てID |
 | エージェントタイプ | CodeWhale、Claude、Codex、OpenClaw、OpenCode |
 | npmパッケージ | [`@misaka-net/fatal-guard`](https://www.npmjs.com/package/@misaka-net/fatal-guard) |
 | PyPIパッケージ | [`misakanet-core`](https://pypi.org/project/misakanet-core/) |

package/README.md

@@ -93,19 +93,24 @@ A **skill** teaches an agent *how to do something*. A **lesson** records *what w
 not to fail again*. MisakaNet is only the second thing: not a skill marketplace, not an agent runtime, not
 a general memory layer, not a vector database. → [FAQ](FAQ.md)
 
-## Benchmark: does lesson context actually help?
+## Benchmark: how much of a lesson does a model reproduce when handed one?
 
-Weekly benchmark on real failure scenarios (Cloudflare Workers AI, 2026-08-30):
+Weekly benchmark (Cloudflare Workers AI, 2026-08-30). **Read the metric before the numbers** — measured
+2026-09-25, the scenario in this benchmark is each lesson's own title, the "matching lesson" injected into the
+`with_lesson` arm is *that same lesson*, and the score is `lesson_hit_rate`: **the share of the injected lesson's commands reproduced in the answer**.
+No retrieval is called and correctness is not checked, so this is the **recitation** half of RAG, not
+evidence that search works:
 
-| Model | Without lesson context | With lesson context | Gain |
+| Model | Lesson pasted in prompt: not pasted | Lesson pasted in prompt: pasted | Difference |
 |---|---|---|---|
-| llama-3.2-3b (light) | 21% hit | **43% hit** | **2× — lesson context doubles a weak model** |
-| llama-3.3-70b (strong) | 42% hit | **73% hit** | **+31%** |
+| llama-3.2-3b (light) | 21% of the lesson's commands reproduced | **43%** | **2×** |
+| llama-3.3-70b (strong) | 42% | **73%** | **+31%** |
 
-Lesson context is a **RAG win across the board**: injecting the matching
-failure-recovery lesson lifts answer quality for every model — the smaller
-the model, the bigger the relative gain. Details:
-[benchmark-2026-08-30](docs/benchmarks/benchmark-2026-08-30.json)
+A model repeats more of a document it was handed, and the weaker the model the bigger the relative
+difference. That is *necessary* for the product to help and it is not sufficient — the claim "search finds the
+right lesson for a failure you described" is measured nowhere yet. Details:
+[benchmark-2026-08-30](docs/benchmarks/benchmark-2026-08-30.json) · metric definition: `METRIC_DEFINITION` in
+[`scripts/benchmark_workers_ai.py`](scripts/benchmark_workers_ai.py)
 
 → [Full changelog](CHANGELOG.md) · [Release notes](https://github.com/Ikalus1988/MisakaNet/releases)
 
@@ -114,8 +119,8 @@ and where this design loses:
 
 | Metric | What it measures | Why it matters here |
 |---|---|---|
-| Hit rate | share of failure questions answered correctly | the only number that decides whether this corpus is worth a search |
-| Gain (with − without) | lift from injecting the *matching* lesson | separates "retrieval works" from "the model got lucky" |
+| Hit rate | share of the **injected** lesson's commands reproduced in the answer — a recitation check; the scenario is that lesson's own title and no retrieval happens | it is the ceiling on usefulness, not the measure of it: a corpus can be recitable and still unfindable |
+| Gain (with − without) | how much more of that lesson appears when it is pasted in | separates "the model can use a lesson" from "the model guessed the same words" — it says nothing about finding the lesson |
 | Cost / latency | tokens and wall-clock per answer | the whole premise is cheaper than re-debugging, so it has to stay cheap |
 
 **Where it loses on purpose:** BM25 matches words, not meaning. A failure described in vocabulary the
@@ -275,8 +280,12 @@ jobs:
 1. Check the checkout works: `python3 scripts/misakanet_cli.py smoke`
 2. Search before writing: `python3 search_knowledge.py "your error here"`
 3. Found nothing? **[Share your failure lesson →](https://github.com/Ikalus1988/MisakaNet/issues/new?template=lesson-feedback.yml)**
-   — a five-line note is enough, no polished PR required. Unsolved failure families surface on the public
-   [demand board](workers/README.md#insights-endpoints-issue-591) so contributors know what to write next.
+   — a five-line note is enough, no polished PR required. Two places say what is missing, and they measure
+   different things: the [demand board](workers/README.md#insights-endpoints-issue-591) aggregates anonymous
+   *search* misses (`/api/insights/unsolved-map`; measured 2026-09-25 it held **one** signal in 30 days,
+   because the intake path does not feed it — [#2224](https://github.com/Ikalus1988/MisakaNet/issues/2224)),
+   and the [open intake issues](https://github.com/Ikalus1988/MisakaNet/issues?q=is%3Aissue+is%3Aopen+label%3Aintake)
+   are where the real gaps currently arrive.
 
 → [CONTRIBUTING.md](CONTRIBUTING.md) · [good first issues](https://github.com/Ikalus1988/MisakaNet/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) ·
 [active competitions](https://github.com/Ikalus1988/MisakaNet/labels/status%3Acompetition) ·

package/README.zh-CN.md

@@ -292,7 +292,7 @@ python3 search_knowledge.py "database locked"
 | 指标 | 数值 |
 |------|------|
 | 📚 Lessons | 411 (canonical, 去重后) |
-| 🌐 Nodes | 1047 |
+| 🌐 Nodes | 2480 |
 | 🎤 Network Voices | 5 条 |
 | 📡 Feed Items | 11 条 |
 | 🔍 领域覆盖 | RAG, DevOps, Feishu, Fanuc, Network, Claude, MCP |

package/SKILL.md

@@ -39,6 +39,20 @@ Do NOT use MisakaNet for:
 
 ## Tools
 
+> **Which surface these examples target.** They are written against the **hosted** endpoint
+> (`https://misakanet.org/mcp`) — the one `docs/mcp.md` recommends and the one most agents are pointed
+> at. The two surfaces differ, and this matters for exactly one tool below:
+>
+> * **hosted** exposes 7 tools, and `misakanet_me_events` is **one of them**
+>   (the stdio set minus `misakanet_submit_usage` / `misakanet_usage_status` /
+>   `misakanet_memory_context`, plus `misakanet_me_events`);
+> * a **local stdio** install (`python3 scripts/mcp_server.py`) exposes 9 tools and **does not have
+>   `misakanet_me_events`** — it has `misakanet_usage_status` and `misakanet_submit_usage` instead.
+>
+> So an agent running the local server will get an unknown-tool error from the reuse-evidence steps and
+> should use `misakanet_usage_status` there. `docs/mcp.md` carries the full table, and
+> `tests/test_mcp_doc_surface.py` keeps this file's tool names in step with it (issue #2000).
+
 ### Register an agent node
 
 ```

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "misakanet",
-  "version": "2.34.0",
+  "version": "2.35.0",
   "license": "Apache-2.0",
   "description": "MisakaNet failure-memory skill — evidence-rated failure lessons for AI agents. DSH bundle: registers the MCP tools as mcp__misakanet__* (misakanet_search/get_lesson/…), served by the public Streamable HTTP endpoint https://misakanet.org/mcp — live tools from an npm install, in any profile, with no local python. A profile that prefers the repo's own stdio server can override the bundle row.",
   "files": [
@@ -44,11 +44,10 @@
     "node": ">=20.0.0"
   },
   "scripts": {
-    "deploy:web": "cd web && npx wrangler deploy",
-    "deploy:api": "cd workers && npx wrangler deploy --config wrangler.api.jsonc",
-    "deploy:email": "cd workers/email-register && npx wrangler deploy",
-    "deploy:all": "npm run deploy:web && npm run deploy:api && npm run deploy:email",
     "deploy": "wrangler deploy",
+    "deploy:api": "cd workers && npx wrangler deploy --config wrangler.toml",
+    "deploy:email": "cd workers/email-register && npx wrangler deploy",
+    "deploy:all": "npm run deploy && npm run deploy:api && npm run deploy:email",
     "preview": "wrangler dev"
   },
   "devDependencies": {

package/skills/misakanet/SKILL.md

@@ -39,6 +39,20 @@ Do NOT use MisakaNet for:
 
 ## Tools
 
+> **Which surface these examples target.** They are written against the **hosted** endpoint
+> (`https://misakanet.org/mcp`) — the one `docs/mcp.md` recommends and the one most agents are pointed
+> at. The two surfaces differ, and this matters for exactly one tool below:
+>
+> * **hosted** exposes 7 tools, and `misakanet_me_events` is **one of them**
+>   (the stdio set minus `misakanet_submit_usage` / `misakanet_usage_status` /
+>   `misakanet_memory_context`, plus `misakanet_me_events`);
+> * a **local stdio** install (`python3 scripts/mcp_server.py`) exposes 9 tools and **does not have
+>   `misakanet_me_events`** — it has `misakanet_usage_status` and `misakanet_submit_usage` instead.
+>
+> So an agent running the local server will get an unknown-tool error from the reuse-evidence steps and
+> should use `misakanet_usage_status` there. `docs/mcp.md` carries the full table, and
+> `tests/test_mcp_doc_surface.py` keeps this file's tool names in step with it (issue #2000).
+
 ### Register an agent node
 
 ```
@@ -49,10 +63,12 @@ Returns a token for authenticated access: it unlocks `misakanet_write_lesson` /
 `misakanet_preflight` and removes the anonymous read limit. **Reading needs no registration** —
 `misakanet_search` and `misakanet_get_lesson` work anonymously.
 
-Pass a `client_id` you can regenerate (a UUID, a workspace id, a hostname) and every later call
+Pass a `client_id` you generate once — **a random UUID you keep private** — and every later call
 returns the **same** `node_id` and token, renewing them; without it each call mints a new node,
-so your reuse evidence, receipts and history start over. `client_id` is an identifier, not a
-credential — the token is always server-issued and random.
+so your reuse evidence, receipts and history start over. Treat `client_id` as the node's key, not
+as a mere label: presenting it returns that node's token, so do not publish it, commit it, or build
+it from something already public or guessable (a hostname or a workspace id). Generate a random
+UUID and store it the way you store a token.
 
 ### Search for lessons