jeroen/cartopolis

Watch

Fork 0

Autopilot log #257

New issue

Closed

opened 2026-09-02 00:24:32 +00:00 by viberfox-agent · 3 comments

viberfox-agent commented

2026-09-02 00:24:32 +00:00

Collaborator

The unattended lane reports here once a day: what it filed, what landed, what
deployed, what is stuck, and how long the login it runs on has left.

It exists because silence is ambiguous — a loop with nothing to do and a loop
that stopped look the same from outside. No new comment for a day means
something is wrong, not that nothing happened.

Nothing reads this issue; close it and it will be made again. The lane itself is
docs/direction.md and the knob is /data/agent/state/autopilot.json.

The unattended lane reports here once a day: what it filed, what landed, what deployed, what is stuck, and how long the login it runs on has left. It exists because silence is ambiguous — a loop with nothing to do and a loop that stopped look the same from outside. **No new comment for a day means something is wrong**, not that nothing happened. Nothing reads this issue; close it and it will be made again. The lane itself is `docs/direction.md` and the knob is `/data/agent/state/autopilot.json`. <sub></sub>

viberfox-agent added the

driven

label

2026-09-02 00:24:32 +00:00

viberfox-agent commented

2026-09-02 00:24:32 +00:00

Author

Collaborator

🤖 2026-09-02

Landed: #255, #251, #249, #245, #243, #241, #238, #230, #218, #217, #216, #210, #206, #203, #201, #199, #197, #192, #187
Filed: #255, #251, #249
Being built: #221, #220
Finished, shipping when CI passes: #237, #219
Stuck, waiting for you: none
main: green on a76cbbe1
Login: 20 days left
Settings: 2/day, 2 at once, shipping off

🤖 **2026-09-02** - **Landed:** #255, #251, #249, #245, #243, #241, #238, #230, #218, #217, #216, #210, #206, #203, #201, #199, #197, #192, #187 - **Filed:** #255, #251, #249 - **Being built:** #221, #220 - **Finished, shipping when CI passes:** #237, #219 - **Stuck, waiting for you:** none - **`main`:** green on `a76cbbe1` - **Login:** 20 days left - **Settings:** 2/day, 2 at once, shipping off

viberfox-agent added this to the Viberfox project

2026-09-02 00:24:52 +00:00

viberfox-agent referenced this issue

2026-09-02 06:54:40 +00:00

gap: a cell's ground waits on api.pdok.nl before it is tessellated #267

viberfox-agent referenced this issue

2026-09-02 06:55:10 +00:00

The daily digest reports Stuck, waiting for you: none while three tickets wait for a person #270

viberfox-agent commented

2026-09-03 00:47:47 +00:00

Author

Collaborator

🤖 2026-09-03

Landed: #276, #269, #266, #263, #258
Filed: #276, #272, #271, #270, #269, #266, #263
Being built: #272, #271, #270, #221, #220
Finished, shipping when CI passes: #237
Stuck, waiting for you: none
main: green on 0b63cd7d
Login: 19 days left
Settings: 1/day, 2 at once, shipping on

🤖 **2026-09-03** - **Landed:** #276, #269, #266, #263, #258 - **Filed:** #276, #272, #271, #270, #269, #266, #263 - **Being built:** #272, #271, #270, #221, #220 - **Finished, shipping when CI passes:** #237 - **Stuck, waiting for you:** none - **`main`:** green on `0b63cd7d` - **Login:** 19 days left - **Settings:** 1/day, 2 at once, shipping on

viberfox-agent referenced this issue

2026-09-03 01:08:27 +00:00

alreadyFiled only ever sees 50 issues, so every autopilot dedupe key silently expires — #206 was QA'd twice #281

viberfox-agent referenced this issue from a commit

2026-09-03 01:44:09 +00:00

fix(autopilot): page every issue list, so a dedupe key does not expire

viberfox-agent referenced this issue

2026-09-03 01:44:37 +00:00

alreadyFiled only ever sees 50 issues, so every autopilot dedupe key silently expires — #206 was QA'd twice #281

viberfox-agent commented

2026-09-10 00:58:22 +00:00

Author

Collaborator

🤖 Retrospective — 2026-09-03 to 2026-09-10 (#296)

The period in five lines.

34 commits landed on main: storeys and stairs behind the BAG's doors, level-crossing masts made solid, a person can upload and share a model, Groningen's own tree, container and charge-point registers, road-surface roughness, and — on 09-09 — the performance suite, the Steam Deck's verdict on the globe cloud shell (71 % of the GPU frame at 1280×720) and a ~50 % cut to it with the floor measured.
The focus lane started: #292 and #294 planned it, and a ten-row queue is in flight. Item 1 (#293, a handheld rung) died on the session time limit; item 2 (#295, cargo perf --repeat N) is refining.
Every build failure in the period was that time limit — #288, #289, #293 — each with "No commits were made". No conflict, no flake.
main went red three times (#278, #290, #291). The fast-forward landing means main only ever moves to a commit CI built, so this is now a branch-quality problem, not a merge-commit one.
On 09-09 the board was swept: 34 qa-gap tickets and nine machinery findings closed, most of them unbuilt — and seven branches carrying finished work were left standing behind them.

What I posted to the queue (#296) — eleven rows, most valuable first:

Page every issue list in the autopilot (the repo holds 112 issues; knownIssues reads 100, so dedupe keys are expiring today). The fix already exists on fix/281-….
The daily note stopped on 2026-09-03 while the lane ticked on 09-05 and 09-09 — and when it posts, mine cannot see a parked ticket it did not file, and nothing names a branch ahead of main.
A parked branch comes back into the lane instead of being addressed to a person.
Rebase and land the three finished map fixes stranded since 2026-08-30 (#214, #232, #268).
A session killed at the 45-minute limit is retried once, then parked as too big.
Decide the three remaining stranded branches — one of them, perf/237, is a performance fix on this lane's own focus.
QA walks docs/qa/targets.md instead of re-verifying every shipped ticket (14 of 14 passes were re-verifies; nine of a documentation-only verdict).
The level-crossing rail gate passes every tile — 64 of 64 Groningen cells, 54 of them wasted (#236).
A re-run BGT extractor must reach a client that already visited the city (#260).
A --frames sweep places every viewpoint, not only the first — the ramp instrument is broken and the fix exists next door in the perf suite.
A surveyed cabinet is not a post box (#259).

What I propose, and why.

The lane's gates end in a person, and the direction says they may not. docs/direction.md states that "every gate has to be a machine one", and three of shipReady's exits — conflict, red branch, no CI verdict — end in a sentence asking a human to rebase. The measured cost of that is eight finished commits sitting on seven branches, two of them for eleven days, with their tickets closed as done. Either the lane gets a machine path back (rows 3 and 4) or the file should say plainly that a parked branch is abandoned, so nobody believes that work landed. I would take the first.

The frontier is no longer the highest-value queue. The frontier lists data sources; the four remaining candidates are each one ticket of checking. Against that, the gap backlog holds measured, cheap, player-facing defects — a rail gate that fetches 54 pointless times per 64 cells, an extractor re-run that reaches nobody who already plays there, a surveyed electricity cabinet that says "Use Post box". #221 argued this in August ("a source is not wired until a player can find it") and was closed unbuilt. My proposal for the file, for you to keep or drop: a source stays on the Wired list only while a player can meet it, and a qa-gap against a wired source outranks the next candidate. That is one paragraph in docs/direction.md, and it is yours to write — I have not touched the file.

The 45-minute limit is now the lane's dominant failure mode, and this repository cannot fix it. The runner that kills the session is not in this tree (tools/ holds only autopilot.ts), so the most useful change — push the worktree before dying, so a killed session leaves its work behind — needs a decision from you about the watcher. Row 5 is the part that can land here: retry once, then park the ticket as too big.

One thing on the focus itself. The queue's item 1 (#293) proposes a CARTO_QUALITY=steamdeck rung by name, and the plan comment on #294 explicitly refused exactly that ("a vendor-family generalisation is not a measurement, and this tree has learned that twice"), while c854c15 landed the measured governor on the same day. I did not re-file it; my read is that it should be dropped in favour of the governor's step 2 (queue item 7), or narrowed to the one part that is not a guess — the Deck takes desktop(), which has no upload ceiling and no residency ceiling at all. Your call, and it is one line in the console either way.

And the Deck is not reachable from the lane. .forgejo/workflows has one runner label, ubuntu-latest, so queue item 6 — the measurement pass that items 7–9 are ordered behind — parks until a person runs it. Worth knowing before the queue reaches it.

_{Filed by the retrospective on #296. No code was changed; no commits were made.}

🤖 **Retrospective — 2026-09-03 to 2026-09-10** (#296) **The period in five lines.** - 34 commits landed on `main`: storeys and stairs behind the BAG's doors, level-crossing masts made solid, a person can upload and share a model, Groningen's own tree, container and charge-point registers, road-surface roughness, and — on 09-09 — the performance suite, the Steam Deck's verdict on the globe cloud shell (71 % of the GPU frame at 1280×720) and a ~50 % cut to it with the floor measured. - The focus lane started: #292 and #294 planned it, and a ten-row queue is in flight. Item 1 (#293, a handheld rung) died on the session time limit; item 2 (#295, `cargo perf --repeat N`) is refining. - Every build failure in the period was that time limit — #288, #289, #293 — each with "No commits were made". No conflict, no flake. - `main` went red three times (#278, #290, #291). The fast-forward landing means `main` only ever moves to a commit CI built, so this is now a branch-quality problem, not a merge-commit one. - On 09-09 the board was swept: 34 `qa-gap` tickets and nine machinery findings closed, most of them unbuilt — and seven branches carrying finished work were left standing behind them. **What I posted to the queue** ([#296](https://code.garage44.eu/jeroen/cartopolis/issues/296#issuecomment-2464)) — eleven rows, most valuable first: 1. Page every issue list in the autopilot (the repo holds **112** issues; `knownIssues` reads 100, so dedupe keys are expiring today). The fix already exists on `fix/281-…`. 2. The daily note stopped on 2026-09-03 while the lane ticked on 09-05 and 09-09 — and when it posts, `mine` cannot see a parked ticket it did not file, and nothing names a branch ahead of `main`. 3. A parked branch comes back into the lane instead of being addressed to a person. 4. Rebase and land the three finished map fixes stranded since 2026-08-30 (#214, #232, #268). 5. A session killed at the 45-minute limit is retried once, then parked as too big. 6. Decide the three remaining stranded branches — one of them, `perf/237`, is a performance fix on this lane's own focus. 7. QA walks `docs/qa/targets.md` instead of re-verifying every shipped ticket (14 of 14 passes were re-verifies; nine of a documentation-only verdict). 8. The level-crossing rail gate passes every tile — 64 of 64 Groningen cells, 54 of them wasted (#236). 9. A re-run BGT extractor must reach a client that already visited the city (#260). 10. A `--frames` sweep places every viewpoint, not only the first — the ramp instrument is broken and the fix exists next door in the perf suite. 11. A surveyed cabinet is not a post box (#259). **What I propose, and why.** *The lane's gates end in a person, and the direction says they may not.* `docs/direction.md` states that "every gate has to be a machine one", and three of `shipReady`'s exits — conflict, red branch, no CI verdict — end in a sentence asking a human to rebase. The measured cost of that is eight finished commits sitting on seven branches, two of them for eleven days, with their tickets closed as done. Either the lane gets a machine path back (rows 3 and 4) or the file should say plainly that a parked branch is abandoned, so nobody believes that work landed. I would take the first. *The frontier is no longer the highest-value queue.* The frontier lists data sources; the four remaining candidates are each one ticket of checking. Against that, the gap backlog holds measured, cheap, player-facing defects — a rail gate that fetches 54 pointless times per 64 cells, an extractor re-run that reaches nobody who already plays there, a surveyed electricity cabinet that says "Use Post box". #221 argued this in August ("a source is not wired until a player can find it") and was closed unbuilt. My proposal for the file, for you to keep or drop: a source stays on the **Wired** list only while a player can meet it, and a `qa-gap` against a wired source outranks the next candidate. That is one paragraph in `docs/direction.md`, and it is yours to write — I have not touched the file. *The 45-minute limit is now the lane's dominant failure mode, and this repository cannot fix it.* The runner that kills the session is not in this tree (`tools/` holds only `autopilot.ts`), so the most useful change — push the worktree before dying, so a killed session leaves its work behind — needs a decision from you about the watcher. Row 5 is the part that can land here: retry once, then park the ticket as too big. *One thing on the focus itself.* The queue's item 1 (#293) proposes a `CARTO_QUALITY=steamdeck` rung by name, and the plan comment on #294 explicitly refused exactly that ("a vendor-family generalisation is not a measurement, and this tree has learned that twice"), while `c854c15` landed the measured governor on the same day. I did not re-file it; my read is that it should be dropped in favour of the governor's step 2 (queue item 7), or narrowed to the one part that is not a guess — the Deck takes `desktop()`, which has no upload ceiling and no residency ceiling at all. Your call, and it is one line in the console either way. *And the Deck is not reachable from the lane.* `.forgejo/workflows` has one runner label, `ubuntu-latest`, so queue item 6 — the measurement pass that items 7–9 are ordered behind — parks until a person runs it. Worth knowing before the queue reaches it. <sub>Filed by the retrospective on #296. No code was changed; no commits were made.</sub>

viberfox-agent referenced this issue

2026-09-10 00:58:45 +00:00

Retrospective: the last 7 day(s), and what to do differently #296