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.
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
opencode writes the session's cwd into the system message's environment
block ("Working directory: …", "Workspace root folder: …"), ahead of the
PR's user message. Each ACP classification ran with its per-call scratch
directory, /tmp/review-classifier-<pid>-<counter>, as cwd, so every call
sent a different system message and the provider's prompt cache could not
match past it.
The session's cwd is now one fixed directory, <tmp>/review-classifier-acp-cwd,
shared by every run. HOME (the config, system.md and opencode's own state)
stays a directory per run, removed after it, as before. With no tool and
no project config nothing writes into the cwd: a live opencode 1.15.10 run
through initialize, session/new and a prompt left it empty, with its state
all under HOME. So concurrent runs can share it, and no state carries from
one run to the next.
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 ACP runtime handed the classifier's system prompt to hive-runtime's
system_prompt_file, which prepends it to the first user message, so the
instructions and the PR text sat in the same user turn.
The rendered opencode.json now defines a review-classifier agent whose
prompt is {file:./system.md} and selects it with default_agent. opencode
1.15.10 sends an agent's prompt as the system-role message (in place of
its per-model default prompt), followed only by its environment block;
the first user message is the PR material alone. The API key runtime
keeps --system-prompt.
The agent carries no permission or tools key, and a test pins its exact
shape: opencode merges an agent's own permission after
OPENCODE_PERMISSION, so a permission there would offer tools again
(checked on real opencode: {"*":"allow"} on the agent offered all ten
tools despite OPENCODE_PERMISSION={"*":"deny"}).
Follow-up to #5214.
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
