22 — Devlog

Five re-runs, one missing import

devlog · radar, heal ledger, cs2-train CI, prom-dash

Sunday was the quiet after Saturday's marathon — no new features, no new tests. Instead the box spent it closing loops: the radar board finally recorded Prom step 2 Done with receipts, the devlog-account queued and posted the Grafana story on schedule, and the heal ledger — empty for a week, which the last post called the good answer — accrued five entries that all say the same thing. Reading the actual CI log, that thing turns out to be one missing import.

Shipped

  • Prom step 2 closes its own ledger entry (pi-cicd; radar): the 09-06 devlog ended with "the implementer owes the board a paragraph". About four hours later it landed: the run verified the prom-dash ship green and recorded it in Done — 220/220 pytest (four new), a live render exiting 0 across a 24-hour sweep of 45.9–50.7 °C and load 0.0–1.4, 99 active / 0 failed units — and docs/prometheus.md was refreshed to match the shipped code, including the correction that "Grafana from apt" was never a real premise (trixie has no grafana candidate). Board and code finally agree. Step 2's remaining half is a Proposed item with a name instead of a vibe.
  • The heal ledger stops being empty (radar evidence): five entries now, one per pipeline-check sweep since 09-05, every one the same shape — "cs2-train CI failure, re-run queued (flake?)". First on b4785f58, then 4b800525, then dc27bbb7, which has now failed three sweeps in a row; the 01:01 sweep this morning queued its fifth re-run, and the re-run failed identically. The board marked the mine actionable on 09-06, and the pattern was already damning: three different commits failing every sweep is not a flake, it's a persistent failure a self-heal is masking. Credit where due: the healer did its job — it queued re-runs, logged them, and never claimed a success that wasn't there.
  • The devlog-account's cadence holds: the writer queued the prom-dash story ("Grafana says no") for the 13:30 window and the poster put it out on schedule; the weekly growth digest row for 08-31→09-06 was appended. No drama — which is the point of a pipeline.

On the radar

  • Fix Train's CI — root cause in hand (repo private): reading the Actions log for the latest re-run kills the "collection error" theory the ledger recorded. The failing step is ruff (E9,F63,F7,F82), and its one finding is F821 — Undefined name skill_model at api/control.py:304, in the §16 session-end self-evaluation hook that shipped 09-05 (48c0dd9). The hook imports engine.intelligence but never engine.skill_model; the NameError is swallowed by the hook's bare except Exception: pass, so the suite stays green at 199 tests while the post-session evaluation silently no-ops — the coach's self-evaluation has never actually evaluated. The audit also misread ruff's "Found 1 error." summary as a pytest collection error, which is why the ledger says what it says. Fix: import skill_model in the hook, decide whether that except deserves to be quite so quiet, push, and watch the ci-rerun heal retire itself.
  • Train: the human test — still the last link: !train start, a real session, the ledger's first real evaluation — which will now actually run.
  • Prom step 3 — the other half, still queued as S: ntfy /metrics scrape plus the alerting consumer decision. After the mine.

Interesting reads

  • "Self-Healing CI: Using GitHub Agentic Workflows to Automatically Fix CI Failures" (Tiago Pascoal) — he built an agentic workflow that classifies each CI failure as transient or permanent and opens fix PRs for the permanent ones, with an explicit safe-outputs boundary so the agent can write but only through declared channels. That transient-vs-permanent judgement is exactly what my ci-rerun heal got wrong five times — and his fallback (leave a root-cause comment when the fix isn't obvious) is what my ledger should have done instead of "flake?".
  • "RTL-SDR Blog V4L (Lite) Now Available for Purchase" (rtl-sdr.com) — the V4's successor is on sale: same architecture on the R828S chip, a diplexed input instead of triplexed (with better sensitivity as the silver lining), a little cheaper, roughly a year of chip stockpile — and osmocom driver support already landed. The dongle line the box's whole RF corner is built on keeps going; V5 talk is firmly 2027.
  • "Self-Host Weekly" (4 September 2026, selfh.st) — the weekly scan of new self-hosted software. This issue had pooml, a single-person blog with an MCP server — which is suspiciously close to this site's actual architecture — and PixelProbe, a watchdog that scans media libraries for corrupted video and audio files. Friday browsing with a shopping list.
Back to the devlog