git:20260920.ad696c9 to git:20260926.4747f47

19 added, 39 removed. Audit A to A.

---

name: lcm-dogfood

disable-model-invocation: true

---

# lcm dogfood

- Run every phase, or only `$0`. The CLI's own help is the checklist: for each command,

- `lcm help <command>` names the options, and the phase passes when every documented option

- ran and did what the help says. Nothing in this file lists options or counts, so it cannot

- go stale; the help can.

- Binary: `lcm` on PATH, or `node dist/bin/lcm.js` after `npm run build`. Large output goes

- through `ctx_execute`; short output through Bash. Record each check as PASS, FAIL or SKIP

- (with the reason), keep going on FAIL, and open an issue for each failure worth tracking.

+ Proves what only a live install can: the installed plugin, the configured summarizer and the user's own memory working together. Hook output shapes, daemon-down behaviour and command options stay out of this skill; `test/e2e/` exercises them against an isolated daemon.

- ## Phases

+ Run every phase, or only `$0`. The target is the real lcm home (`LCM_HOME`, default `~/.lossless-claude`):

+ - Phases that write (`compact`, `sensitive`, `mcp`) touch only the current project, or add and then remove their own entries. A run costs one project's summarization and keeps every existing summary; `--all` and `--restart` stay out of this skill.

+ - Phases that only read (`health`, `integrity`) cover the whole home.

- **health**: `lcm status`, `lcm doctor`, `lcm --version`. Done when the daemon is up on this

- project, every doctor check passes or its warning is recorded, and the version equals

- `package.json`.

+ Binary: `lcm` on PATH. Record each check as PASS, FAIL or SKIP (with the reason), keep going on FAIL, and open an issue for each failure worth tracking. The daemon log is `logs/daemon.log` in the lcm home, one JSON record per line; `docs/configuration.md` ("Daemon log") names its events.

- **import**: `lcm help import`, then every option. Done when a second identical run adds no

- message beyond the current session's own.

+ ## Phases

- **compact**: `lcm help compact`, then every option. The summarizer is an LLM call; allow

- five minutes. Done when a second identical run creates nothing.

+ **health**: `lcm --version`, `lcm status`, `lcm doctor`. Done when the version equals `package.json`, the daemon is up on this project, and every doctor check passes or its warning is recorded. The `daemon-log` check covers errors and continuity in the daemon log.

- **promote**: `lcm help promote`, then `lcm stats --verbose`. Done when the promoted count

- moved, or the output states there was nothing promotable, and the stats agree with the

- counts the earlier phases printed.

+ **capture**: the plugin's hooks reaching the daemon from this session. To find this session's id, echo a unique marker (such as `dogfood-` plus the epoch seconds) in one command, then, in the next, search this project's directory under `~/.claude/projects/` recursively for `*.jsonl` files containing it. The one file that holds it is this session's transcript, and its name without `.jsonl` is the session id. Done when the daemon log holds a `request` record for `POST /prompt-search` with `status` 200, `cwd` equal to this project and `session_id` equal to this session's id.

- **sensitive**: `lcm help sensitive`, then every subcommand. Done when a built-in secret and

- a pattern you added are both `[REDACTED]` by `lcm sensitive test`, and the pattern you added

- is gone from `lcm sensitive list` after you removed it.

+ **import**: `lcm import --dry-run --verbose`. Done when it lists this project's sessions and `lcm status` shows the same message count before and after.

- **hooks**: `docs/hook-protocol.md` is the contract. Done when every hook command it lists

- answers a valid stdin payload with the output shape it documents, within its timeout, and

- the daemon's `/prompt-search` answers `.claude/skills/lcm-dogfood/scripts/prompt-search-test.js <query>` directly.

+ **compact**: note the time and the `Summaries` and `Promoted` counts `lcm status` shows for this project, then `lcm compact`, which compacts the current project and then promotes. The summarizer is an LLM call; allow five minutes. Done when:

+ - a second identical run creates nothing;

+ - either `lcm compact` printed that there was nothing to compact (it then sends no request, so the log has no record), or the daemon log holds a `compact.done` or `compact.skipped` record for this project written after the noted time;

+ - every `summarizer.fallback` record from that interval is noted in the scorecard with its `reason`;

+ - `lcm status` afterwards shows this project's `Summaries` count higher than before when the log holds a `compact.done` for this run, and unchanged otherwise; `Promoted` is never lower than before. `lcm stats` totals every project, so it cannot confirm one project's counts.

- **mcp**: `docs/agent-tools.md` is the contract. Done when every tool it lists has been

- called with a documented parameter set and answered in the documented shape, including a

- `lcm_store` followed by an `lcm_search` that finds it. Skip `lcm_expand` and `lcm_describe`

- only when no summary id exists yet.

+ **sensitive**: `list`, `add`, `test` and `remove`, from `lcm help sensitive`. It writes the real pattern list, so remove every pattern you add. `purge` deletes stored memory, so this skill never runs it. Done when a built-in secret and a pattern you added are both `[REDACTED]` by `lcm sensitive test`, and your pattern is gone from `lcm sensitive list`.

- **resilience**: stop the daemon (`lcm daemon stop`), then `lcm status`, a hook command, and

- `lcm daemon start --detach`. Done when the down state is reported without a hang, the hook

- returns within its timeout with empty or valid output, and status shows the daemon up

- again.

+ **mcp**: `docs/agent-tools.md` is the contract. Done when every tool it lists has been called with a documented parameter set and answered in the documented shape, including an `lcm_store` followed by an `lcm_search` that finds it. Skip `lcm_expand` and `lcm_describe` only when no summary id exists yet.

- **debug**: `~/.lossless-claude/daemon.log` tail, `.claude/skills/lcm-dogfood/scripts/db-integrity.js`, and `$PWD`

- against `pwd`. Done when no ERROR line is unexplained, every project database reports

- `ok`, and the two paths match.

+ **integrity**: `node .claude/skills/lcm-dogfood/scripts/db-integrity.js`, which checks every project database read-only; on a large home it takes minutes. Done when it prints `N of N project databases ok`.

## Scorecard

- One row per phase run: checks, PASS, FAIL, SKIP. For each FAIL: the error, the daemon log

- excerpt, the suggested fix. Done when every phase you ran has a row.

+ One row per phase run: checks, PASS, FAIL, SKIP. For each FAIL: the error, the daemon log records around it, the suggested fix. Done when every phase you ran has a row.