Watch
0
0
Fork
You've already forked hyperhive
0
hyperhive/swarm-review-classifier/README.md
atlas ce7b3e055f review-classifier: ask a check once more when its answer isn't valid JSON
An answer that fails to validate is asked for a second time, for that
classification only, with a reminder to reply with the JSON object alone.
The second ask comes out of the classification's CLASSIFICATION_DEADLINE:
it gets what the first ask left, and is skipped when less than 20 seconds
remain, so RUN_DEADLINE's bound still holds. A failed ask (runtime error,
timeout) is not repeated. If the second answer does not validate either,
the row has no answer, as before.

Model::ask takes the deadline for the ask; the ACP runtime no longer holds
its own.

The parser also takes the first fenced block of a reply that has prose
around the fence; prose around a bare object is still rejected.
2026-10-16 16:55:07 +02:00

36 lines
2 KiB
Markdown

# swarm-review-classifier
The review classifier's try-out service, in the `swarm-review-classifier`
container on the host that sets `deploy.review-classifier.enable`. The
design is [`docs/swarm/review-classifier.md`](../docs/swarm/review-classifier.md).
swarm-controller reads a PR (its head commit, its commits and its diff) and
sends it, with the operator's classifications, as a request on the queue.
This service asks a small model one classification at a time and answers
with the result table and the overall verdict. It writes nothing to the
forge and holds no forge credential.
## What a run is
- The shared context first: a system prompt, the PR's commits with their full
messages, and the diff with lockfiles collapsed as `hive-forge diff` does.
Every classification's prompt starts with the same bytes, and `claude` gets
that part as a content block of its own, with a cache breakpoint at its end.
- One `claude --print` per classification, with no session, no tools
(`--tools ""`), no MCP server, its own empty directory and
`ANTHROPIC_API_KEY` in its environment only.
- An answer is a JSON object with a verdict, a certainty and a reason; the
prompt asks for a reason under 350 characters, and the classifier truncates one over the 500 cap to 499 plus `…`. A failed run, a
timeout or an answer that doesn't validate leaves that row without a verdict, and it doesn't count.
An answer that doesn't validate is asked for once more first, with a
reminder to reply with the JSON object alone, within the same three minutes.
- A run with no API key, or with no classification answered, fails as a
whole: the caller sees an error, never a verdict.
## Credentials
The container holds one credential: its own store certificate, which nspawn
hands in at start without a copy on disk. With it the service
reads its queue client secret at start and the API key at every run, so a key
written or removed in the store takes effect on the next run. Its store role
reads exactly those two paths.