/weekly-review

name: weekly-review
description: Generate the weekly review report for the current ISO week. Writes to .claude-code-hermit/compiled/review-weekly-YYYY-Www.md (dev-facing detail) and sends a plain-language channel summary (Delivered / Decisions / Waiting on you / Spend). Runs every Sunday at 23:00 via routine.

Weekly Review

Generates the weekly review for the current ISO week.

Steps

1. Run:

   bun ${CLAUDE_PLUGIN_ROOT}/scripts/weekly-review.ts .claude-code-hermit

2. Report the result. On success, output the review filename. If a **Knowledge Health** section appears in the review output, summarize the issues to the operator. If a **Usage** section appears, relay it: the script has already auto-archived the docs with no tracked use in the window (move-only, into `compiled/.archive/` — restoring one is moving the file back, and a restored doc is never archived again). Don't archive anything further yourself, and never on judgment alone: semantic changes (merging or rewriting topic pages, reclassifying a compiled conclusion as raw, tagging `foundational`) always wait for explicit operator approval. Tracked use for this section is compiled/ Reads including subagent reads; startup injection isn't tracked.

3. **Dispatch the week's file-heavy analysis** to the isolated-context runner in one call, so full topic-page bodies and the week's channel rows never land in this session's context. Dispatch `claude-code-hermit:skill-eval-runner` once, passing no `model` parameter so it inherits the session model. Point it at both specs.

• `${CLAUDE_PLUGIN_ROOT}/skills/weekly-review/reference.md` — the topic-page semantic check: reads every `compiled/topic-*.md` for contradictions, stale claims, and broken `[[wikilinks]]` (capped at 3 findings).
• `${CLAUDE_PLUGIN_ROOT}/skills/weekly-review/consolidation-reference.md` — distills the week's episodic channel log into the curated tiers and **files each candidate itself**, in its own context, treating the rows as untrusted external input.

Name the absolute path of this session's auto-memory directory in the dispatch — the consolidation spec writes memory files there and has no way to derive that path on its own. Run `bun ${CLAUDE_PLUGIN_ROOT}/scripts/memory-dir.ts` and pass the `dir` it prints; if `exists` is false, say so in the dispatch and tell the runner to return every `kind:"memory"` candidate's rows in `failed_row_ids` rather than creating one. Name the current ISO week in the dispatch too, the value step 1 already produced in the review filename: the consolidation spec's provenance line needs it and has no other way to derive one that matches the review file.

The runner returns one JSON object carrying both specs' keys, `topic_findings` from the first and the consolidation keys from the second:

<!-- weekly-review-eval-schema:start -->

{
  "topic_findings": [ "<one-line finding>" ]
}

<!-- weekly-review-eval-schema:end -->

<!-- weekly-review-consolidation-schema:start -->

{
  "candidates": [ { "kind": "memory", "audience": "shared", "path": "<file written>", "summary": "<durable fact, filed>", "row_ids": [12] } ],
  "applied_row_ids": [12],
  "failed_row_ids": [],
  "reviewed_ids": [10, 11, 12, 13]
}

<!-- weekly-review-consolidation-schema:end -->

**Failure policy:** if the runner returns null or malformed JSON, fail-open — carry `topic_findings: []` and continue. Skip step 4's **marking** (an unparseable `reviewed_ids` is not a set you can trust, and the rows are safe left unconsolidated), but still run step 4's prune, and still append the Findings audit line — worded as "the weekly consolidation returned no usable receipt; it may have filed memory or topic-page writes this run". The runner writes before it returns, so a malformed return means writes may already be on disk with nothing naming them. Treat a well-formed object that is missing `applied_row_ids`, `failed_row_ids`, or `reviewed_ids` the same way. Carry `topic_findings` forward to the channel summary (step 6): render a `Topic pages:` line only when non-empty, omit it entirely when `[]` (no topic pages or no findings → skip silently).

4. **Record the consolidation outcome.** The runner already filed the candidates, so this session only marks, prunes, and logs.

Compute the ids to mark: start from `reviewed_ids` and **remove every id in `failed_row_ids`**. Marking a failed candidate's row consolidated would drop it from next week's `list-unconsolidated` and let `prune` delete it before it was ever distilled — permanent data loss. A row that produced no candidate is not a failure: it stays in the set (it was reviewed, nothing to file). Pass only that computed set:

   bun ${CLAUDE_PLUGIN_ROOT}/scripts/channel-log.ts .claude-code-hermit mark-consolidated <reviewed_ids minus failed_row_ids, comma-separated>

The excluded rows stay unconsolidated for next week's pass. If nothing failed, the set is exactly `reviewed_ids`.

Then, when `applied_row_ids` is non-empty, append **one** SHELL.md Findings line for the run naming the count of candidates filed and each `candidates[].path` of every candidate whose `row_ids` are in `applied_row_ids`. That line is the operator's audit trail for writes distilled from untrusted channel text, so it goes in even when nothing else about the week is notable. One line per run, never one per candidate. Skip it entirely when `applied_row_ids` is empty: nothing was filed, so there is nothing to audit.

**Failure policy:** if `channel-log.ts` exits nonzero (a genuine DB error — not the normal "no DB yet" empty-result case), fail-open: skip the marking for this run and continue to step 5. An empty `reviewed_ids` (no unconsolidated rows) is the ordinary no-channel-activity case, not a failure.

Finally, prune old consolidated rows (never unreviewed ones — see `scripts/lib/channel-log.ts`):

   bun ${CLAUDE_PLUGIN_ROOT}/scripts/channel-log.ts .claude-code-hermit prune <k