pmf-review

name	pmf-review
description	Roll up the week's daily PMF logs into a Weekly Review entry (Focus / Learned / Decisions / Next), surface recommended hypothesis signal/status updates for you to confirm, and refresh the index. Use at the end of a week, or when the user says 'pmf review', 'do my weekly review', 'roll up the week', 'weekly synthesis', or types '/pmf-review'.
argument-hint	[optional: week-start YYYY-MM-DD, and/or path to the pmf/ root]
version	0.1.0
user_invocable	true

/pmf-review — weekly rollup

You are synthesizing a week of daily logs into one honest weekly review. Roll up, don't inflate: three n=1 notes on the same point may become n=3, but only say so explicitly and only if justified — never launder repetition into a trend.

Before you do anything

1. Read ${CLAUDE_PLUGIN_ROOT}/references/honesty-rules.md (NON-NEGOTIABLE), format-spec.md, and evidence-weight.md.
2. Resolve the tracker root $PMF and the week window: default to the last 7 days ending today (date +%Y-%m-%d); if $ARGUMENTS gives a week-start date, use [start, start+6d].

How to run it

There are two equivalent paths. Prefer the workflow if your environment runs workflow scripts; otherwise do it inline. Both produce the same review.

Path A — workflow (typed, parallel)

The canonical implementation is ${CLAUDE_PLUGIN_ROOT}/workflows/pmf-review.js. It runs four phases — Collect (resolve root + window, list the week's logs, read the prior review's ## Next), Read (one sub-agent per daily log, in parallel, extracting did/learned/convos/movement/decisions with evidence weights preserved — no cross-day synthesis), Synthesize (one agent drafts the four review fields and justifies any evidence-weight rollup), and Write (write the review file, record recommended signal/status changes for you to confirm via /pmf-wrap, regenerate the index). It never edits hypothesis files itself. Invoke it as a workflow if available.

Path B — inline (always works)

1. Collect. List $PMF/log/*.md whose date is in the window. Read the newest $PMF/reviews/*.md's ## Next.
2. Read. Read each in-window log. For many logs, fan out one read-only sub-agent per log (Task/Agent tool) to extract {what_did, what_learned, customer_convos, hypothesis_movement, decisions} verbatim with evidence weights — instruct each sub-agent to follow honesty-rules.md and NOT synthesize across days.
3. Synthesize. From the extracts + the prior ## Next, draft the four fields:
   • ## Focus — what the week was actually about.
   • ## Learned — grounded findings; roll up evidence weight explicitly (show the n=), no invented trends.
   • ## Decisions — decisions actually made.
   • ## Next — concrete next steps (carry forward unfinished items from the prior ## Next).
4. Write. Create $PMF/reviews/<week-start>.md from ${CLAUDE_PLUGIN_ROOT}/templates/weekly-review.md; set period, logs_rolled_up (the filenames), and hypotheses_touched (union across the week). Record any hypothesis signal or status changes the week suggests as recommendations in the review's ## Next (e.g. "Recommend H4 signal — → Mixed; confirm via /pmf-wrap"). Per honesty-rules.md rule 4, neither signal nor status changes to hypothesis files happen here without explicit confirmation — apply them in /pmf-wrap, not in the review. (If you want to apply a change now, switch to /pmf-wrap, which gates it with AskUserQuestion.)

After writing

Regenerate the index: python3 ${CLAUDE_PLUGIN_ROOT}/scripts/pmf_index.py --out "$PMF" --check. Report the review path, which hypotheses changed, and that the index refreshed. If the week had zero logs, say so and write nothing.