monitor-projects

# Monitor Projects Nightly, on a scheduler cron, review **the skills amended during the day** — not the whole agent-skills project (agent-skills is a collection of skills; the unit of review is the skill directory) — via a five-seat concurrent roundtable, then store the result in `/memory` for later recall. ## Pipeline (one `nightly` run) 1. **Discover** — fetches, then selects skills whose `skills/<name>/` paths were touched since the **last reviewed commit** (`<watermark>..HEAD`). The watermark lives at `~/.local/state/monitor-projects/watermark.json` (override with `MONITOR_PROJECTS_WATERMARK`) and advances **only after a run that both succeeded and stored** — so a failed or missed night re-scans the same range instead of losing it. `--since` is the first-run fallback only; a wall-clock window cannot resume, and a missed night would silently skip everything in the gap while its receipt still looked complete. An unknown watermark falls back to the window rather than selecting nothing. No amendments → `no_changes` receipt, exit 0. 2. **Context** — build the shared packet from: - `/project-state report --json --cached` (project readiness evidence), - `/ops-workstation` quick health (host context), - `/brave-search web` for each amended skill's load-bearing topic (capped, external evidence per `/best-practices-roundtable`). 3. **Roundtable** — one `/ask` compile+execute, per `/best-practices-roundtable` (equal context, concurrent topology, no privileged seat): ```bash skills/ask/run.sh tau-dag "<shared packet>" \ --repo local/agent-skills --target monitor-projects-<date> \ --handler webgpt --handler webclaude --handler webkimi \ --handler webgrok --handler webgemini \ --topology concurrent --execute --poll-timeout-seconds 3600 --json ``` 4. **Synthesize** — per-seat status (responded / blocked / stale tab / timed out), common ground, attributed dissent, executable slices. A missing seat is `NEEDS_ATTENTION`, never silent consensus. 5. **Store** — via the memory daemon only (NEVER direct ArangoDB): - full receipt → `POST /store` `collection: project_roundtables` (searchable via `/memory recall`; the memory repo registers this collection in the ArangoSearch view per `arango-recall-all-collections`); - compact summary → `POST /store` to `lessons` (default collection) with tags `["monitor-projects", "roundtable", <date>, <skills...>]`, which gets Qdrant semantic sync + dedup, guaranteeing hybrid recall. - **Read-back verification**: the run only reports `stored` after `/memory recall` returns the document. A `/store` 200 is not proof. ## Commands ```bash cd skills/monitor-projects ./run.sh discover --json # amended skills in last 24h (no side effects) ./run.sh nightly # full pipeline (discover→context→roundtable→store) ./run.sh nightly --dry-run # everything except --execute and /store ./run.sh last # recall the most recent roundtable from /memory ./run.sh discuss "<question>" # recall roundtable receipts relevant to a question ./run.sh register # register the nightly cron with /scheduler ``` ## Cron `./run.sh register` runs: ```bash skills/scheduler/run.sh register \ --name monitor-projects-nightly \ --cron "30 2 * * *" \ --command "skills/monitor-projects/run.sh nightly" ``` 02:30 sits before the 03:00–04:15 monitor-taxonomy/monitor-skills window. ## Retrieval and discussion Both the human and the project agent retrieve results the same way: ```bash ./run.sh last # newest receipt, rendered skills/memory/run.sh recall --q "monitor-projects roundtable <topic>" --brief ``` Every stored document carries `schema: monitor_projects.roundtable.v1`, `date`, `skills_reviewed`, `seat_status`, `common_ground`, `attributed_dissent`, `executable_slices`, and the ask run directory path so receipts can be re-read in full. ## Proof boundaries - Seat responses are **advisory reviewer evidence**, not local proof. Executable slices still require deterministic local verification by the project agent before closure. - `nightly --dry-run` proves discovery, context assembly, and packet shape, and never advances the watermark (asserted by `sanity.sh`). It does not prove live browser transport; only an `--execute` run with validated receipts (`validate_live_browser_workflow.py`) proves that. ## Eval posture `sanity.sh` runs behavioral gates over committed fixtures: positive control (amended-skill discovery on a synthetic git repo), watermark controls (range selects the right commit, a watermark at HEAD selects nothing, an unknown watermark falls back), negative control (no amendments → `no_changes`), packet-shape assertion (every seat receives the identical packet; all five handlers present), and safety boundary (dry-run performs no `/store` and no `--execute`). Live roundtable transport is covered by /ask's own release gate, not duplicated here. ## Common Mistakes ### WRONG: reviewing agent-skills as one project The unit is the amended skill directory, `skills/<name>/`. ### WRONG: writing to ArangoDB or Qdrant directly ```python from arango import ArangoClient # forbidden ``` ### RIGHT: memory daemon only ```python client.post("http://127.0.0.1:8601/store", json={"document": {...}, "collection": "project_roundtables"}) ``` ### WRONG: trusting /store's 200 response ### RIGHT: read back via `/memory recall` before reporting `stored`. ### WRONG: giving one seat extra context because its transport is easier ### RIGHT: one shared packet, identical for all five seats, concurrent topology.