rule-card.json ---> PreToolUse hook

     |                          ^                          |
     v                          |                          v
user correction  ---- Stop hook capture                 BLOCKED
     |                                                     |
     +-----------> proof console <-----------+             |
                        |                   |             |
                        v                   v             |
                evidence/ceo-summary.html <--+-------------+
                        |
                        v
                        CEO

-->

metrix-plugin

Previous Claude Code made this mistake.
New Claude Code tried to repeat it.
TeamAgent blocked it.

A Claude Code plugin marketplace that captures user corrections in one session and prevents the same mistake from repeating in the next.

_{Real interactive claudefast session in tmux — two distinct Claude Code sessions, real model round-trip, real PreToolUse denial rendered by the TUI. 72 s end to end (idle compressed). Cast file: teamagent-demo.cast. Earlier mock recording archived at docs/demo/archive/.}

---

What is this

Built on the TeamBrain idea: short-term agent memory becomes durable team policy.

┌─────────────────────────────────────────────────────────────────┐
│  Alice corrects Claude    → teamagent-memory captures rule card │
│  Bob retries 3 days later → teamagent-memory blocks via hook    │
│  CEO opens the dashboard  → teamagent-proof-console renders HTML│
│  Team-wide propagation    → teamagent-team-sync ships the rule  │
│  Third-party verdict      → bin/judge.sh + jq + ffprobe say PASS│
└─────────────────────────────────────────────────────────────────┘

Install

/plugin marketplace add LiuShiyuMath/metrix-plugin
/plugin install teamagent-memory@metrix-plugin
/plugin install teamagent-proof-console@metrix-plugin
/plugin install teamagent-team-sync@metrix-plugin

The hooks register themselves via each plugin's hooks/hooks.json. The rule store lives at ~/.teamagent/rules.jsonl; the event log at ~/.teamagent/events.jsonl.

Three plugins

Plugin	Category	Job
teamagent-memory	memory	Capture corrections via Stop hook; block repeats via PreToolUse; inject rule context via UserPromptSubmit.
teamagent-proof-console	observability	Render evidence/ceo-summary.html (4 verbatim anchor strings + before/after diff) from real captured events. /teamagent-proof-console:proof
teamagent-team-sync	collaboration	SessionStart pulls team rules + writes conflicts.jsonl instead of overwriting; UserPromptSubmit suggests publish when the user wants to share.

Verify

bin/judge.sh is a third-party harness — bash + jq + ffprobe + node --check + claudefast. The LLM never self-judges; it can only read judge.json afterwards, in mechanical isolation via --plugin-dir /tmp/empty.

bash bin/judge.sh           # runs three probes, writes judge.json, prints PASS / FAIL
bash bin/judge-verdict.sh   # ships judge.json to a blank-brain LLM judge under --plugin-dir /tmp/empty

Three probes:

Probe	What it proves
stream-json	claudefast actually emits a parseable stream-json event source
ab-plugin-dir	A: empty plugin dir → no block. B: rule seeded → block with permissionDecision: deny. C: benign command → still passes. Causal A/B + selectivity.
file-checks	rule-card.json schema · ceo-summary.html ≥ 2 KB + 4/4 anchors · node --check for every hook · git status --porcelain clean · optional ffprobe on demo mp4

Demo it yourself

The recording you see at the top of this README was generated by:

bash demo/teamagent-demo.sh                                 # raw run, real hooks, isolated HOME
asciinema rec docs/demo/teamagent-demo.cast \
  --cols 120 --rows 36 \
  --command "bash demo/teamagent-demo.sh" --overwrite       # record
agg docs/demo/teamagent-demo.cast docs/demo/teamagent-demo.gif \
  --font-size 14 --speed 1.2 --idle-time-limit 1.5          # cast → gif

Try it locally — the script lives at demo/teamagent-demo.sh and exercises every hook for real (no LLM round-trip needed).

Docs

• docs/duck-guidebook/index.html — single-page Chinese guidebook for a non-coder CEO reader (鸭语 throughout, ten chapters, opens in a browser, zero build steps).
• docs/demo/teamagent-wild.gif — real-machine 4-pane tmux recording with separate HOME / CLAUDE_HOME lanes.
• demo/teamagent-sandbox-server.cjs — local clickable browser console that controls a real tmux sandbox; Alice is user-controlled and Leader watches changes.
• EVAL.md — the tool-enforced eval contract. claudefast --plugin-dir /tmp/empty does the final read-only judging.
• plugins/*/README.md — per-plugin user docs (hook contracts, file layouts, troubleshooting).
• probes/README.md — judge probe specs and how to add a new one.

Repo layout

metrix-plugin/
├── .claude-plugin/marketplace.json     ← catalog (3 plugins, version 0.1.0)
├── plugins/
│   ├── teamagent-memory/               ← 3 skills, 3 hook .cjs, bin/teamagent
│   ├── teamagent-proof-console/        ← 3 skills, /proof slash command
│   └── teamagent-team-sync/            ← 3 skills, 2 hook .cjs
├── evidence/                           ← rule-card.json, ceo-summary.html
├── bin/                                ← judge.sh, judge-verdict.sh
├── probes/                             ← stream-json.sh, ab-plugin-dir.sh, file-checks.sh
├── demo/                               ← teamagent-demo.sh (deterministic end-to-end)
├── docs/
│   ├── demo/                           ← teamagent-demo.cast + .gif
│   └── duck-guidebook/index.html       ← CEO-duck single-page guidebook
├── README.md                           ← (you are here)
└── EVAL.md                             ← third-party verdict contract

License

MIT — see each plugin's manifest. Built by @LiuShiyuMath, based on the TeamBrain marketplace concept.