The wide diff's context drops from 200 to 25 lines, so a block's
enclosing gate stays in view when it opens within 25 lines above the
change (the forge's own diff carries 3). The fallback note, the doc and
the gate fixture (now 20 lines out, not 47) follow.
A failure reading the changed files (a forge error on a tree or blob, a
truncated tree, a non-base64 blob) or joining the diff task made the PR
read fail, so the review posted no verdict. It now falls back to the
forge's 3-line .diff with "reading the files failed: …" as the note.
render wrote each file's whole diff before wide compared it with
MAX_BYTES. It now writes through a writer that refuses once past the cap,
which stops similar's formatter mid-file, so one huge file costs about
MAX_BYTES of output, not its full diff.
The classifier read the forge's PR `.diff`, whose context is fixed at 3
lines. A block moved into a glue module lost its enclosing gate (a
`config = lib.mkIf …` 13 to 191 lines up in the misfired moves #5199,
#5200, #5201), so the glue restating that gate read as a new condition.
The controller now builds the diff itself: it lists the PR's changed files,
reads each at the merge base and at the head through the git tree and blob
API with the token it already holds, and diffs them with 200 lines of
context. It sends the forge's `.diff` instead when the PR changes more than
100 files, the forge names no merge base, or the wide diff is over 256 KiB,
and the posted review then says so with the reason.
Fixes#5227
The retries covered about a minute, and the swarm-bao restart that lost
#5286's review lasted 3m19s. Each try now gets its own forge deadline,
a try with no answer by then is retried like a refused connection, and
the waits are 5s, 15s, 45s, 90s and 150s: six tries over about five
minutes.
The docs also said to request the review again after a post-failed run.
That does nothing while the account's earlier request is still pending,
so they now say to remove the request and then request it again.
A review request's verdict was lost when the post failed: on 2026-10-14
the classifier approved #5286, but its post hit swarm-bao mid-restart
(cert login: connection refused), the run was recorded post-failed, and
nothing on the PR showed it.
post() now tries the post up to four times, waiting 5s, 15s and 45s,
when it fails because the connection to the secret store or the forge
failed or timed out. Any other error fails at once, as before. The
verdict is computed once and reposted as is. All tries share the one
forge deadline the single post had, and the warn log carries the number
of tries.
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.
The two review tabs, runs and try-out, are one: /review. It is a column of
three sections: try a classification and the classifier settings, each a
folded bar (the settings' bar says what they are now), then the recent runs,
always open. The settings had been pinned under the try-out form; they are now
a section of their own. A new Disclosure primitive is the folding bar, a
native details wearing the Panel's classes, so the folded and open sections
read as the same surface.
/review-runs and /review-tryout are gone, with no redirect.
The commit author's email was the forge account's own address, read with
the site-admin token, which put an email the account may keep private
into the target repo's history. It is now the address Forgejo itself
uses for such an account's web commits, `<lowercased login>@<NO_REPLY_ADDRESS>`
(models/user/user.go GetPlaceholderEmail). NO_REPLY_ADDRESS defaults to
`noreply.<DOMAIN>` (modules/setting/service.go) and the forge module
leaves it unset, with DOMAIN set from services.hyperhive.swarm.forge.domain,
the same option SWARM_CONTROLLER_FORGE_URL is built from; so the domain is
that URL's host. The account is still read, to refuse a user with no
forge account, but its email no longer is.
Refs #5259
The try-out page gets an "open a PR" button. It posts the page's
classifications to the new POST /api/review-classifier/open-pr, which
merges them by name into the repo's rules file on the default branch
and has the review-classifier forge account commit the result on a new
branch, review-classifier/<user>-<8 hex>, through the contents API. The
commit's author is the forge account named like the session's
Remote-User, read with the controller's admin token. The answer links
to the forge's compare page, where the user opens the PR as themselves.
The push uses the account's existing token from the store
(swarm/services/review-classifier/forge/token, scope write:repository),
so the account needs write access to the repo; a refused push answers
403 naming that. In a YAML rules file, the entries the merge leaves
alone keep their text, as long as the result reads back as the merged
list; otherwise the file is written out whole.
DefaultBranchFiles now carries each file's blob sha, which the forge
needs to replace a file, and the repo's web URL for the link.
Fixes#5259
Refs #5096
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
An answer whose reason ran past 500 characters was rejected whole, so the
verdict was lost and the one error row turned the review into no_verdict.
The reason is cut to the cap, ending in an ellipsis; verdict and certainty
are kept.
A run record this build couldn't decode was dropped from the list it
read, and the next recorded run wrote the list back without it, so a
rollback followed by a roll-forward lost the newer build's runs for
good. The stored list is now a list of entries, each a decoded run or
the raw JSON of one that didn't decode. The route still shows only the
decoded runs. An undecodable entry keeps its place, counts toward the
100, and is written back as it was read. A read that finds any logs one
WARN with their count and the first error.
The settings decode derived its known fields from what
`Settings::default()` serializes, so a later field skipped when unset
would have been stripped from a valid stored document and lost on the
next PUT. The known fields are now an explicit list. A test serializes
a fully populated `Settings`, checks that its keys match the list, and
checks that `decode` round-trips it.
The controller keeps the classifier's newest runs as one JSON array and
its settings as one JSON object, each under one KV key. A document this
build could not decode, as after a rollback past a build that added an
outcome or a settings field, failed every read and every write until
someone deleted the key by hand. A classifier newer than its controller
refused every run request for the missing `model`.
The run store now decodes the array one record at a time: a record that
does not decode is dropped with a WARN naming the bucket, key, index and
error, and a document that is not an array reads as no runs. Listing
keeps working, and the next recorded run writes back only the records
that decoded.
The stored settings ignore a field this build does not know; one that
still does not decode reads as the defaults with a WARN. The PUT body
keeps `deny_unknown_fields`. A PUT now always writes, so a PUT of the
defaults also replaces an undecodable document.
A run request without `model` runs on haiku, the swarm-wide setting's
default, now one constant in swarm-queue-client.
`schedules::read` gains a sibling, `read_with`, taking the decoder.
Closes#5162
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.
Every run is recorded when it ends: a try-out when its reply settles it,
a requested review once its review is posted or the post fails, and a
request turned away by a full queue once its comment is posted. A
duplicate request that starts nothing is not recorded. The newest 100
are one JSON array under `recent` in a new controller-owned KV bucket,
`review-classifier-runs`, written with compare-and-set as
`schedule-definitions` is (five attempts). A failed write is logged; the
review is already posted by then.
Each recorded run is also one info event, "review classifier run
finished", with the record's fields as tracing fields. The controller
logs natively to the journal, so they reach VictoriaLogs as F_<NAME>.
GET /api/review-classifier/runs returns the list, newest first, gated
by the gateway like the try-out GET. post_classifier_review now returns
the posted review's html_url, which the record keeps as review_url.
Closes#5153, part of #5096.
Two read-only routes feed them: GET /api/pulls/repos lists the repos with an
open pull request, and GET /api/pulls?repo=owner/name lists a repo's open
pull requests, highest number first. The page offers the lists through the
browser's own suggestion list and keeps both fields free text, since a run
also works on a merged or closed PR.
Follow-ups from the review of #5142.
A `Done` answer with fewer or more rows than the classifications sent, or
none, now posts the no-verdict comment instead of the classifier's overall
verdict, so a short answer can't approve.
Requests are accepted per repo, PR and head into a set bounded at the two
running plus eight waiting. A request for a key already running or waiting
starts nothing and logs at info; a request with every place taken starts no
run and posts a no-verdict comment saying so. Before, every delivery spawned
a task waiting on the semaphore, so toggling the review request queued one
paid run per toggle.
Table cells get a zero-width space after `#`, `!`, `://` and `www.`, so a
model's reason can't render an issue or PR reference or a bare URL as a link.
The secret publisher's policy denies the classifier's forge token leaf, the
way #5139 denied Grafana's admin password, with a module-eval-bao-grants arm
pinning it. The controller posts reviews with whatever token is there, so a
publisher write picked the account reviews come from.
Closes#5143, part of #5096.
Requesting the forge account `review-classifier` as a reviewer now runs the
repo's rules against the PR and posts the result as that account's review,
on the head commit the run read. A push starts no run; only a fresh review
request does.
- An instance-wide `pull_request_review_request` system hook delivers to
`/webhook/forge/review-request`. A request for any other reviewer, or a
removed request, does nothing.
- The controller reads the PR, loads the rules with the same reader as
`GET /api/review-classifier/rules`, sends one request on the try-out
subject, and posts APPROVED or REQUEST_CHANGES with the table as the body.
- With no clean verdict (PR or rules don't load, the run fails, or any
classification has no answer) it posts a COMMENT saying why, never an
approval. The event is one function, `no_clean_verdict_event`.
- Table cells are flattened, markup-escaped, and every `@` gets a
zero-width space, so a model's reason can't break the table or mention
anyone.
- At most two requested reviews run at once; the rest wait in memory.
- The account is minted the way hive accounts are: a five-minute pass
queues a `MintClassifierForgeToken` node when the token at
`swarm/services/review-classifier/forge/token` is missing or stale. The
token is `write:repository`. The controller's bao policy gains that leaf
(create/read/update), pinned in module-eval bao-grants. The classifier
never sees the token.
The account gets no repo access automatically; a repo admin adds it by
hand. A re-request retires the account's earlier review: Forgejo clears
`official` on a reviewer's earlier reviews when it is requested again
(models/issues/review.go AddReviewRequest, v16.0.5), and approvals and
rejection blocks count official reviews only.
Refs #5137
forgejo-api escapes the whole file path of a raw-file request, so the
rules file went out as `raw/.hyperhive%2Freview-classifier.yml` and
reached the forge only if every proxy and router in between decoded the
`%2F`. The controller now sends that one request itself: `raw_file_url`
escapes each path segment on its own and joins them with `/`, and
`forge::Client` holds a second HTTP client with the same token for it.
A 404 from the default-branch lookup now means "no commit on the
default branch", like an empty repo, instead of falling through to the
route's "no repo" answer. Only the repo lookup's 404 still means the
repo doesn't exist.
Tests cover the URL shape, the reads against a loopback stub forge
that serves only the literal path, the branch 404, 503 with no forge
access, 503 at the forge deadline on a paused clock, and a bad
polarity value.
Part of #5096.
`GET /api/review-classifier/rules?repo=owner/name` reads
`.hyperhive/review-classifier.yml` or `.json` from the repo's default
branch with the controller's forge token, on every request, and answers
the classifications in the try-out POST's shape, plus the file, the
branch and the commit it was read at. The swarm UI's try-out page loads
them into its editor (#5134).
The file is one `classifications` list of the try-out's `Classification`.
A missing file, both files, a parse error, any unknown key, an empty list
or more than ten entries is a problem+json error, never a pass. The
checks the POST makes on its classifications move into one function both
paths call.
The doc names the file and its schema, replacing the placeholders, and
documents the route.
Part of #5096.
- review-classifier.md: claude runs spawn through hive-claude directly
with no tools and no session (hive-runtime holds only TurnStop); the
answer's fields, the 500-character reason cap and what a failure does;
how the try-out page reaches the classifier; the settled parts of the
open list (API key path, try-out subject, --tools "" on claude).
The banner stays.
- services.md: a Review classifier section with the two operator steps.
- secrets.md, credentials.md: the classifier's leaf and the key row.
- network.md: the container's row. CLAUDE.md: the crate.
Add docs/swarm/review-classifier.md, the high-level architecture of the
planned PR review classifier pipeline, written as the full target shape
under a "planned, not implemented yet" banner. Link it from docs/README.md
and the docs/swarm/README.md piece list.
The page reflects the epic description and the operator's answers in the
epic thread; points still open are listed in its "Still open" section.
The banner names the epic by title rather than by number because the
tracker-tag lint rejects issue references in tracked files.
