Skip to content
noggin

Retro: the 2026-11-01 → 2026-11-02 session (noggin-87)

Sources: this session’s transcript and its subagents’ reports, issues and PRs #334–#481 with their comments, the messages exchanged with the parallel session (noggin-27), CLAUDE.md, CONTRIBUTING.md, .claude/agents/implementer.md, the Makefile and .github/workflows/ on main. Ranked by what each cost. Decisions are Cansin’s (grilling, 2026-11-02).

# Finding Kind Fix
1 Claude’s review skipped silently when a PR’s review.yml differed from main’s (claude-code-action’s workflow validation) and reported only “the review produced no verdict”: #390 twice, #395. Each cost a diagnosis round. Automated check #433 / #436: Review rebases such a branch itself as the release App and fails naming the cause; the App’s token is minted only on that path
2 Agents pushed without running the checks and learnt from CI: pyright (#391), the vocabulary test’s v2 ban (#388), ruff line length (#409). make check builds every part in CadQuery first, so agents skipped it. Automated check #427 / #454: make check-light (~20 s, no parts build) and a .githooks/pre-push hook that make venv turns on; implementer.md names it as the only local check
3 PR titles were wrong at creation (a scope used as the type, scopes not allowed, no scope): #378, #339, #349, #400, #408. Automated check + steering #388 / #399: strict type(scope) titles, scopes derived from the repo; #430 / #434: implementer.md checks the title with scripts/pr_title.py first
4 Two sessions rebased the same PR (#388, #390, #394); one push went out mid-rebase (#394 briefly lost a commit) and one past a failing lint (#409). Tool economy #429: scripts/rebase_pr.sh (refuses mid-rebase, runs check-light, pushes with lease) and a session:<name> PR label
5 Both sessions named worktrees ../noggin-<n>, so one rebased the other’s (#390); ~60 stale worktrees piled up. Navigation #428 / #437: worktrees under ../noggin-wt/<session>/<branch>, make worktrees-clean
6 make worktrees-clean’s first run removed an open PR’s worktree (#460): a pushed branch whose tip main already held counted as merged. Automated check #464 / #466: never remove a branch with an open PR; “done” means its PR is merged or closed
7 Claude decided “approve” but never posted it (“the approval didn’t land”, 4+ times): the structured verdict ended the session before gh pr review ran. Automated check #389 / #391: a hook refuses the verdict until the matching review is on the head
8 Agents stopped and resumed on every CI event, and one reported #391 as waiting on Cansin after it had merged. Tool economy #430 / #434: one blocking wait that ends on a terminal state; re-read the PR before reporting
9 Homebrew uv 0.12.19 couldn’t install the pinned Python 3.14.8; make venv failed with “No interpreter found”. Information access #431 / #435: [tool.uv] required-version; uv upgraded on the Mac
10 Main’s own checks went red twice and blocked every PR: a stale generated README table (#423), then zizmor unpinned-images after #462 + #463 (#476 / #477). Automated check Fixed by #423 and #477; the README table is gone since #425, and #481 drops the pins

Fixed during the session besides: PySpice → spicelib (#337), the parked MQTT broker deleted (#336), Python 3.14 (#344), ADR 0012 (keep everything current, the Core’s firmware through the vendor’s tools), the eyes Core’s facts (#349: LittleFS, USB-C data one way round, firmware 1.1.4.13), the Mech’s version in its beacon, flashed and checked live (#366).

Information access: the eyes Core shows no USB device in KEY download mode, so its flash encryption stays community evidence (docs/build/flashing.md). The sandbox hides /dev/cu.*; reading the Core needed the sandbox lifted.

No-ops: CLAUDE.md gained one directive line (keep everything current); nothing in it went unused this session.

byYellow PineYellow Pine

noggin v0.2.26-6-g09bf9b7 · 09bf9b7