* fix: align reviewer eval with published findings (#1713)
* fix: make reviewer eval reflect published findings
Serialize and deduplicate finding persistence, align review calibration around the final six-finding publication, and make judge matching order-independent and auditable.
* fix: honor reviewer eval limits
Forward configured caps into publication snapshots and keep recall-at-cap bounded for diagnostic all-findings runs.
(cherry picked from commit 71e3b8183882bcc42e318f3f220c291617ebcb67)
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* Empty commit to trigger CI
---------
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* Run reviewer eval in a GitHub Action; make dashboard a read-only progress view
The dashboard launched the eval as a subprocess inside the serving deployment
worker, so a container recycle killed long runs and discarded results that had
already completed server-side. Move the harness to a workflow_dispatch Action
(run on prod). run_eval now publishes status/progress/log-tail to the LangGraph
store record the dashboard reads, so /admin/evals stays a live view; a killed
Action surfaces as failed via the stale-heartbeat reconcile.
* reviewer_eval workflow: pass inputs via env, no shell interpolation
Addresses the reviewer finding: workflow_dispatch string inputs were
interpolated into the run: block (limit unquoted), allowing shell injection in
a job holding LANGSMITH/ANTHROPIC keys. Pass inputs through env and reference
quoted "$VARS"; validate limit is numeric and build its flag in bash.
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Drive dashboard-triggered reviewer eval runs with per-run model, effort,
score mode, severity threshold, cap, limit, and concurrency overrides, plus
per-example start/finish/error logging in the eval target.
Rework the admin eval form from the label-left/control-right SettingsRow
(which crushed the description column when packing 3-4 wide inputs) into
stacked field groups with captioned inputs in a responsive grid.
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: trigger reviewer evals from the admin page
Add an admin-only "Reviewer eval" section + endpoints that launch the
reviewer benchmark as an isolated subprocess against the running
deployment, with live status and the LangSmith experiment link. Route
eval traces to a dedicated open-swe-evals project so they stay out of
the production tracing project.
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* fix: reconcile reviewer eval status via heartbeat, not local process
The persisted record is shared across workers but _PROCS is process-local.
The owning worker now refreshes a heartbeat while the subprocess runs, and
status is only reconciled to failed once the heartbeat is stale, so a poll on
a worker without the local handle no longer kills a live run (and a duplicate
start is rejected across workers).
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
evals/reviewer/{target,run_eval,build_dataset} called load_dotenv() at
import time, so importing them in tests injected the real .env (live
LANGGRAPH_URL, tokens) into the whole pytest process. The slack-context
default-repo tests then reached the real LangGraph store via
get_team_default_repo() and picked up the developer's actual team
default repo, failing in full-suite runs while passing in isolation.
Move load_dotenv() into the CLI entrypoints (all env reads were already
lazy), and patch get_team_default_repo in the two affected tests so they
stay hermetic regardless of environment.
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: tune reviewer for precision — web/wiki tools + recalibrated prompt
Reviewer agent now has web_search, fetch_url, and http_request alongside the
finding tools, so it can verify library semantics and consult the DeepWiki
auto-generated wiki for public repos (https://deepwiki.com/<owner>/<repo>)
before flagging cross-file or architectural concerns.
Prompt rewritten to push precision over recall:
- explicit severity ladder pushing reviews toward bimodal high/low instead of
defaulting to medium
- ≤200-char description target (gold set averages ~186 chars; we were at ~436)
- mandatory docs / wiki / code lookup before flagging concurrency, security,
or perf — the three categories that dominated false positives
- "do not flag" list covering compiler/linter-catchable nits, speculative
claims without a concrete attacker/interleaving/scale, style preferences
the codebase doesn't share, and test-quality nits on non-test diffs
- smart file-selection guidance for large PRs (deprioritize generated /
vendored / pure-rename hunks)
Eval config switched to openai:gpt-5.5 + high reasoning effort for the next
benchmark run.
* trim prompt
* subagent prompting
* confidence ratings
* added medium
* enforce confidence threshold
* .
* reviewer: precision-tuned prompt + drop confidence gate
Rewrites the reviewer system prompt around a defensibility bar (anchor +
failure mode + maintainer wouldn't say "not a bug"), an explicit do-not-file
list (style nits, speculation, scope-policing, same-bug fan-out), and a
checklist of 10 bug archetypes drawn from a per-PR audit of the eval golden
set. The audit showed 145 FPs in the last eval split ~28% speculative, ~26%
style-nit, ~31% real-but-unscored (mostly same-archetype fan-out); the new
prompt targets each class directly.
Confidence is still recorded on every finding for post-hoc calibration but
no longer gates publication — the audit showed the gate was a no-op (agent
self-rated 65% of findings "high" regardless), and the prompt's defensibility
bar is the actual discipline. Drops CONFIDENCE_ORDER, CONFIDENCE_THRESHOLD,
the confidence_threshold kwarg on filter_findings_for_publish, the
confidence_filtered score_mode, and the min_confidence kwarg on the eval
target's _extract_comments — all dead once the gate is gone.
Also removes the "informational" severity tier from the Severity enum,
SEVERITY_ORDER, and all validators / tests / docstrings. It was reserved for
FYI observations the dataset never rewards.
* benchmax
* adding google provider
* slight steering
* tuning
* more tuning
* fix
* cleanup
* reducing overfitting
* Add per-repo review style profiles and inject them into the reviewer.
Dashboard users can analyze historical PR review feedback per repository,
edit the resulting style guide, and have it loaded from LangGraph Store at
reviewer runtime (including Martian eval runs) keyed by owner/name.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review style job errors leaking exception details to clients.
Return generic dashboard messages while logging full stack traces server-side.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat: tighten reviewer eval workflow
Require the reviewer to verify and dedupe findings before recording them, and make benchmark runs safe to execute against deployed reviewer graphs without posting GitHub reviews.
* chore: move reviewer eval settings to config
Load reviewer benchmark settings from the default eval config file so deployed eval runs do not require a wide CLI surface.
* feat: allow reviewer eval model overrides
Pass reviewer model and reasoning effort from the eval config into reviewer runs so isolated benchmark deployments can test Opus 4.7 high thinking.
* fix: use adaptive thinking for Opus 4.7
Switch Opus 4.7 model overrides to Anthropic adaptive thinking with effort instead of the deprecated budgeted thinking payload rejected by the API.
* refactor: use latest Anthropic effort API
Remove legacy Anthropic budget-token thinking support and route Anthropic efforts through adaptive thinking plus effort.
* revert prompting
* feat: add reviewer graph + eval target wiring
- New `reviewer` graph (`agent/reviewer.py`) registered in langgraph.json
alongside the main `agent` graph. Reuses the same sandbox lifecycle,
GH proxy auth, and middleware primitives from `agent.server`, but with
a narrower tool set, a reviewer-specific system prompt, no
commit/push, and the `task` (subagent) tool stripped via
`_ToolExclusionMiddleware` so review stays in one context.
- New `github_comment` tool: agents call it once per issue with
`(file, line, body, severity)` and the eval scores those calls
against golden comments.
- `ensure_no_empty_msg` middleware (the no_op nudge) is intentionally
*not* on the reviewer's stack — that middleware exists to enforce the
main agent's "always finalize via Slack/Linear/PR" contract, which
the reviewer doesn't have. The main agent's behavior is unchanged.
- `evals/reviewer/target.py`: send PR info as a user message, extract
every `github_comment` tool call (multiple expected per review) into
the run output.
- `evals/reviewer/judge.py`: per-example evaluator now returns a list
of metrics under `{"results": [...]}` so LangSmith averages each
numeric key (f1/precision/recall/tp/fp/fn) across the experiment in
the UI. Dropped the broken `aggregate_pr` summary evaluator that
reached for an attribute that doesn't exist on `RunTree`.
- `evals/reviewer/run_eval.py`: `--limit` now slices the dataset via
`client.list_examples(limit=N)` since `aevaluate` doesn't accept
`max_examples`.
- Makefile: `dev` and `run` targets now use `uv run` so they work
without an activated venv.
* resolve comments
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: add reviewer eval harness
Offline LangSmith eval scaffolding for the upcoming Open SWE Reviewer graph.
Imports the 50 PRs from withmartian/code-review-benchmark goldens, resolves
base/head SHAs via gh, and runs a claude-opus-4-5 LLM judge using martian's
verbatim prompt so scores are directly comparable to their published Devin
Review numbers. Reviewer graph itself is not part of this change.
* fix: ruff lint and format on reviewer eval files
