Part of #72. Follows #168. Today the router only ranks cards someone wrote, so optimization work shows up only when a human notices it. Astra should find it on its own, deterministically.
Changes (router.py and astra.py):
Where it looks, each source a synthetic candidate with key opt:<source>:<id>, kind optimization, and a drafted title and body:
The ledger, last 7 days (time/ledger/<YYYY-MM>.jsonl): spend by model and by entry point. If one route is over half the week's spend, or a model's spend per call is double its 7-day median, that's a candidate ("Why did sonnet cost $X on Y?").
Actions runs, last 7 days: any workflow with 2 or more failed or cancelled runs (/actions/runs, as tea_actions reads them) is a candidate, with the failure count and the last run id.
Parked cards (error / stuck beats) older than a day with no open PR: one candidate each ("retry or split #N"), unless #78's auto-retry already has them.
How it's ranked: the seat judges these like any candidate.
In live mode, a picked one becomes a written card (unqueued), with the drafted body and a DONE: line. It is never worked on in place.
In dry mode, the table shows it.
No duplicates: a source with an open card that names it (Router: opt:<source>:<id> in the body) is not proposed again.
Tests: a ledger with one dominant route yields one candidate. Two failed runs of one workflow yield one. A candidate with an open card yields none. Live writes the card; dry doesn't.
DONE: PR open with the code above, python3 .github/time/test_time.py green including the new checks, CI green
Part of #72. Follows #168. Today the router only ranks cards someone wrote, so optimization work shows up only when a human notices it. Astra should find it on its own, deterministically.
Changes (`router.py` and `astra.py`):
- **Where it looks**, each source a synthetic candidate with key `opt:<source>:<id>`, kind optimization, and a drafted title and body:
- The ledger, last 7 days (`time/ledger/<YYYY-MM>.jsonl`): spend by model and by entry point. If one route is over half the week's spend, or a model's spend per call is double its 7-day median, that's a candidate ("Why did sonnet cost $X on Y?").
- Actions runs, last 7 days: any workflow with 2 or more failed or cancelled runs (`/actions/runs`, as `tea_actions` reads them) is a candidate, with the failure count and the last run id.
- Parked cards (`error` / `stuck` beats) older than a day with no open PR: one candidate each ("retry or split #N"), unless #78's auto-retry already has them.
- **How it's ranked:** the seat judges these like any candidate.
- In live mode, a picked one becomes a written card (unqueued), with the drafted body and a `DONE:` line. It is never worked on in place.
- In dry mode, the table shows it.
- **No duplicates:** a source with an open card that names it (`Router: opt:<source>:<id>` in the body) is not proposed again.
- **Tests:** a ledger with one dominant route yields one candidate. Two failed runs of one workflow yield one. A candidate with an open card yields none. Live writes the card; dry doesn't.
DONE: PR open with the code above, `python3 .github/time/test_time.py` green including the new checks, CI green
⏱️ Time: adopted PR #199, built outside the worker. haiku reviewed it on the PR.
⏱️ Time: adopted PR #199, built outside the worker. `haiku` reviewed it on the PR.
<!-- time {"beat":"pr","pr":199,"adopted":true,"launched":false,"seat":"haiku","as":"seat","review":"venice","seats":["haiku"]} -->
⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment /time retry to send it back to the worker with the review attached).
⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment `/time retry` to send it back to the worker with the review attached).
<!-- time {"beat":"needs-human","pr":199} -->
☁️ The cloud session pushed a revision to PR #199. The heartbeat reviews it again; a verdict from before the push no longer counts.
☁️ The cloud session pushed a revision to PR #199. The heartbeat reviews it again; a verdict from before the push no longer counts.
<!-- time {"beat":"dispatched","pr":199,"revised":true,"cloud":true} -->
⏱️ Time: PR #199 is up for this card. haiku reviewed it on the PR.
⏱️ Time: PR #199 is up for this card. `haiku` reviewed it on the PR.
<!-- time {"beat":"pr","pr":199,"launched":false,"seat":"haiku","as":"seat","review":"venice","seats":["haiku"]} -->
⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment /time retry to send it back to the worker with the review attached).
⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment `/time retry` to send it back to the worker with the review attached).
<!-- time {"beat":"needs-human","pr":199} -->
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Part of #72. Follows #168. Today the router only ranks cards someone wrote, so optimization work shows up only when a human notices it. Astra should find it on its own, deterministically.
Changes (
router.pyandastra.py):opt:<source>:<id>, kind optimization, and a drafted title and body:time/ledger/<YYYY-MM>.jsonl): spend by model and by entry point. If one route is over half the week's spend, or a model's spend per call is double its 7-day median, that's a candidate ("Why did sonnet cost $X on Y?")./actions/runs, astea_actionsreads them) is a candidate, with the failure count and the last run id.error/stuckbeats) older than a day with no open PR: one candidate each ("retry or split #N"), unless #78's auto-retry already has them.DONE:line. It is never worked on in place.Router: opt:<source>:<id>in the body) is not proposed again.DONE: PR open with the code above,
python3 .github/time/test_time.pygreen including the new checks, CI green☁️ The cloud session's branch came home as PR #199. The heartbeat reviews and lands it like any card PR.
⏱️ Time: adopted PR #199, built outside the worker.
haikureviewed it on the PR.⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment
/time retryto send it back to the worker with the review attached).☁️ The cloud session pushed a revision to PR #199. The heartbeat reviews it again; a verdict from before the push no longer counts.
⏱️ Time: PR #199 is up for this card.
haikureviewed it on the PR.⏱️ Time: the reviewer objected on PR #199. Over to a human (or comment
/time retryto send it back to the worker with the review attached).