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.
The review body's table is now: result | check | certainty / threshold |
reason. The result cell is a symbol for what the row did to the review,
after polarity (✅ counted and passed, ❌ counted and failed, ❔ not
counted or no answer), then the model's verdict. The overall line carries
✅ / ❌. Polarity, threshold and counted are folded into those cells.
The prompt asks for a reason under 350 characters; the enforced cap
stays 500, so an overshoot is truncated less often.
On the API key runtime the classifier now hands claude its prompt as one
stream-json user message (`--input-format stream-json`) with two text
blocks: the PR's shared context, carrying `cache_control: ephemeral`, and
the classification's question.
As plain stdin text the whole prompt was one block, and claude puts its
only message breakpoint on the last block, after the question. The cache
writes an entry only where a breakpoint is, so no entry ended at the
shared-context/question boundary. Every classification wrote the whole
prompt and none read it (production: cacheRead ~0, cacheCreation ~input).
The system-prompt prefix (~270 tokens) is under the minimum cacheable
length. Captures against a local listener with a dummy key (claude
2.1.223) show the client breakpoint survives on the shared block, and that
claude adds none there on its own.
The PR stays in the user role: moving it into the system prompt would give
untrusted PR text the system prompt's authority.
`Model::ask` takes a `Prompt { shared, question }`. The ACP runtime joins
it back into the same text it sent before.
Refs #5245
Every classification's shared context began with the PR's head sha,
headed each commit with its sha, and carried each file's
`index <hash>..<hash>` line in the diff. A push or rebase rewrites all
of those, so the provider's prompt cache missed from the first line of
the user message even when the commit messages and hunks were
unchanged (#5245: "cache hits show as 0.1x").
The header now names only repo#pr, each commit is headed by its
position, and the diff drops `index` lines before lockfiles collapse.
The `diff --git` paths and every hunk stay. The TryOut's head_sha is
untouched: the controller still posts the review against it, and the
journal line still logs it.
On the ACP runtime the cost sink also adds each classification's tokens
to `hyperhive.review_classifier.token.usage`, typed input / output /
cacheRead / cacheCreation like claude's own `claude_code.token.usage`,
so the hit rate is a number in the metrics store on both runtimes.
The swarm settings gain `runtime` (`api-key`, the default, or `acp`) and
an `acp` provider record: an https base URL, the provider's model, and
its prices per million tokens. A PUT now changes only the fields its
body names, so the settings form that sends model, effort and limit
leaves the runtime alone; `null` clears `effort` or `acp`. The body stays
strict, and the stored document keeps #5192's lenient read.
The controller sends the runtime in each run request. On `acp` the
classifier runs one prompt on a new `opencode acp` session per
classification through hive-runtime's AcpRuntime, in a scratch
directory, with a config rendered from the settings. opencode runs
`read` and other tools without asking, so OPENCODE_PERMISSION
{"*":"deny"} takes every tool away from the model, and a deny-all policy
refuses any request on top. The title call, which resends the prompt,
is turned off. The deadline cancels the turn, and a turn that does not
end within 15s of that is dropped, which also covers an agent stuck
before its prompt. The answer is the final assistant text, as ACP has
no result event.
The provider key is read per run from
swarm/services/review-classifier/acp-provider, which the bao UI's
admins login may create and update, and reaches opencode through
AcpRuntime::with_provider_key and the trampoline's allowlist.
Each ACP classification's cost is computed from the configured prices
and the reported tokens, and added to
hyperhive.review_classifier.cost.usage labelled repo, pr, model and
currency.
Part of #5096
The classifier ran every classification on a hard-coded `haiku` with no
`--effort`, and the controller ran at most two requested reviews at once,
also hard-coded, so changing any of them took a code change and a
redeploy.
swarm-controller now keeps one settings document,
`{model, effort, max_concurrent_reviews}`, in its own
`review-classifier-settings` KV bucket under `all`, the way it keeps the
schedule definitions. No document means today's values: haiku, no
`--effort`, two reviews. `GET /api/review-classifier/settings` reads it;
`PUT` replaces it whole, needs the gateway's `Remote-User` like the
try-out POST, and refuses a bad model (the inference route's check), an
effort outside the five levels or a limit outside 1..=8 with
problem+json 400.
Every run request (try-out and requested review) now carries the model
and effort, and the classifier passes them to claude instead of its
constant, which is gone. A try-out reads the settings when it is posted;
a review reads them for the limit before it waits for a place and again
for the model and effort once it has one. The review semaphore is
replaced by a running count under a resizable limit, so a PUT, or the
next review's read, changes how many run without a restart, and a lower
limit lets the runs already going finish. MAX_WAITING_REVIEWS is
unchanged.
Closes#5151, part of #5096.
A new service crate. It reads its queue client secret from the store at
start, takes try-outs on the review classifier subject in its queue
group, and for each run reads the Anthropic API key from the store,
builds the shared context (commits with full messages, diff with
lockfiles collapsed), runs one claude --print per classification with no
session, no tools, no MCP server and the key in the child's environment
only, validates the JSON answer and aggregates the rows. A missing key or
a run where no classification got an answer fails the run; one failed
classification is an excluded row with its error. One journal line per
run names the operator, repo, PR, head SHA, rows and overall verdict.
swarm-secret-client gains the two paths the service reads.
