Workflow Retro

One question: what friction repeated this week, and what single change removes it?

Query via ywai. Raw DB reads only for the freshness check in references/queries.md.

Rules

1. Fix scope first: days (default 7), project or worktree. No scope, no retro.
2. Collect every source before concluding. One failed source never stops the run.
3. Validate freshness before analyzing. A blind tracker is finding #1, not a footnote.
4. Cluster only on >=3 evidences. One anecdote is not a finding.
5. Falsify before proposing. Each hypothesis states what would disprove it; check that first.
6. One friction = one lever (skill|agent|tool|code) + one experiment. Never two changes in one experiment. Classify first: mechanical (fixed syntactic pattern, banned API, import shape, file-location rule) goes to a deterministic check (lever code|tool: repo linter rule, pre-commit hook, or CI job); judgement call (cross-file consistency, matches surrounding style, needs intent) goes to skill|agent. Default to building the check. Place the fix where it runs. Review imposes standards. Implementation follows pointers. AGENTS.md holds pointers only.
7. Report as HTML (see references/report.md). OS temp dir, open it, report the absolute path. Nothing lands in the repo.
8. mem_save only the top 3 as pattern. The detail lives in the report.

Process

1. Scope

Ask at most: project/worktree + days. Default days=7. Record the exact filters in the report header.

2. Collect

Run in order, --json always. Full command list in references/queries.md:

• Freshness: session_v2 vs session counts + max dates, raw SQL, read-only.
• ywai eval run --days <n> --worktree <w> --json — sessions aggregate.
• GET /api/evals/runs + /api/evals/summary — hard scores, turns, invalid, cost, attempt errors.
• GET /api/engram/search?q=<symptom> + /prompts + /observations + POST /api/engram/memory-evals — recall quality.
• git log --since=<n>.days — oneline + name-only + authors.
• Guardrail: repo check command + hooks + CI workflow (lint/check scripts, pre-commit hook, CI lint/typecheck/test job). Read before proposing a new check. An unwired or broken check is the finding, not a reinvention.

3. Validate

Apply references/thresholds.md: thin window → widen AND retitle; stale tracker → finding #1; thin memory → Speculative. Pick ONE primary window here; everything downstream uses it.

4. Cluster

Rules first (references/thresholds.md), LLM second. Method in references/analysis.md: rework done right, early-death vs hard-task, cost-vs-overhead, cross-source joins.

5. Interrogate

For each cluster, write the disproof before the proposal: the one check that would kill your hypothesis (attempt error field, revert list, author split). Run it. Dead hypotheses die here, silently.

6. Propose

Survivors become cards: evidence (3 sessions + metrics) / hypothesis / lever / ONE experiment / success metric. Rank by frequency × impact × fixability. Fewer Strong findings beats padded ones. A repo with no guardrail (no pre-commit hook and no CI job running lint/typecheck/test) is itself a finding, not a neutral default.

7. Report + save

Render assets/report-template.html with the data. Buttons for PDF + PNG ship in the template. Then mem_save top 3.