* feat: upgrade default agent + reviewer model to Opus 4.8
Replace Opus 4.7 with Opus 4.8 (claude-opus-4-8) as the supported
Anthropic model surfaced in the profile editor and used by the main
agent and reviewer graphs. Effort levels (low/medium/high/xhigh/max)
and the high default are unchanged, matching the official Opus 4.8
docs. Updates eval config comment and tests accordingly.
* fix: provider-aware fallback for stale stored model ids
Dropping claude-opus-4-7 from the supported set meant persisted
profile/team-settings still holding it failed the SUPPORTED_MODEL_IDS
check and fell through to default_model_pair() — a cross-provider jump
to the OpenAI global default.
Add provider_fallback_pair: when a stored id is no longer supported but
its provider still has a supported model, resolve to that provider's
newest supported model (anthropic:claude-opus-4-7 -> 4.8), preserving
effort when valid. Resolution order is now: valid stored pair ->
same-provider fallback -> global default_model_pair(). Profile overrides
keep deferring to the team default when no model is set or the provider
is unknown.
* feat: reconcile reviewer findings with PR threads
* fix: harden reviewer finding reply handling
* fix: queue reviewer finding reply body
Ensure review-comment replies that arrive during an active reviewer run include the sanitized reply body in the queued reassessment prompt.
* fix: apply reviewer reply formatting
Apply the repository formatter so the reviewer reply handling fix passes CI format checks.
* reviewer: fetch PR diff via GitHub API to re-enable add_finding validation
The previous hotfix in reviewer.py set diff_line_set=None because the
sandbox-based diff prep was sometimes producing empty diffs. That made
every bad anchor a publish-time 422 instead of a creation-time
rejection — the agent burned tokens producing unanchorable findings,
and we had to add a publish-time retry safety net (#1338) to clean up.
Fetch the PR's unified diff via the GitHub REST API at reviewer
startup and populate diff_text + diff_line_set so add_finding can
reject bad anchors immediately. The API path is reliable and is the
same diff GitHub validates against when posting inline review
comments. If the fetch fails, fall back to the previous behavior
(validation disabled, publish-time retry handles it).
Also extract the PR-diff fetch into reviewer_diff.fetch_pr_diff so
both reviewer.py and publish_review.py share one implementation
instead of two copies.
* reviewer: make diff_line_set validation side-aware
compute_diff_line_set previously returned only new-side line numbers,
so re-enabling add_finding's validation would wrongly reject findings
with side=LEFT (deleted-line bugs whose only anchor is an old-side
line). Return {file: {"RIGHT": {new_lines}, "LEFT": {old_lines}}}
instead, and have is_range_in_diff select the matching side from the
finding's recorded side. add_finding and publish_review's retry
filter both pass the finding's side through.
* fix(reviewer): inject existing PR review threads into reviewer context
The reviewer agent was filing the same inline comment on every re-review
because it only saw findings recorded on its own thread metadata — not
the live PR review-thread state on GitHub. When a previous finding was
still open (code unchanged, or a human reply explained it), the agent
rediscovered the same defect on the next push and called `add_finding`
again, producing duplicate comments.
This change fetches the PR's review threads (across all reviewers, with
replies and isResolved status) via GraphQL and renders them into the
first-review and re-review contexts as a "Pre-existing PR review
threads" block. The system prompt now lists overlap with that block as
a hard "Do NOT file" rule, and treats threads addressed by a human
reply as resolved.
This also gives the reviewer comment-awareness on its very first run
on a PR, so it skips findings already raised by another reviewer or
bot.
* fix(reviewer): wrap PR review threads in untrusted-data XML block
Addresses the reviewer comment on this PR
(https://github.com/langchain-ai/open-swe/pull/1331#discussion_r3295497533):
PR review comment bodies are attacker-controlled (anyone who can comment
on the PR can put anything in them), and they were being concatenated
into the reviewer's system prompt with instruction-priority.
Switches the existing-threads section from a Markdown block to an XML
data block:
<pr_review_threads>
<thread location="path:line" status="open">
<comment author="open-swe[bot]">
<body>...</body>
</comment>
<comment author="romain-priour-lc">
<body>We added defaults in the template</body>
</comment>
</thread>
</pr_review_threads>
The system prompt now explicitly names the wrapper, tells the agent that
everything inside it is untrusted data from the PR (not instructions),
and that prompt-injection payloads inside bodies must be disregarded.
We keep the bodies so the agent can actually read engineer replies —
that's the whole point of comment-awareness — but they're delimited as
data, not concatenated as prose. Modern frontier models are well-trained
to honor this contract.
Additional defenses:
- Author logins are validated against the GitHub username grammar; any
unexpected value is rendered as "unknown" so the `author` attribute
can't smuggle freeform text.
- Literal closing tags (`</body>`, `</pr_review_threads>`, etc.) in
bodies are neutered so a body can't break out of its wrapper.
- Body length is capped at 4000 chars per comment to bound the prompt.
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat(reviewer): inline AGENTS.md into reviewer system prompt
Fetches AGENTS.md from the PR's head_sha via the GitHub contents API
during reviewer setup and inlines it as a "Repository conventions"
block in the system prompt. Mirrors how the per-repo review style
prompt is already wired.
The main agent has long had a mandatory step to read AGENTS.md after
cloning, but the reviewer often skips cloning entirely (it can
`gh pr diff` directly), so it never saw the file. Loading it
deterministically means the reviewer judges findings against the
project's own conventions instead of relying on the model to fetch
the file itself.
* fix(reviewer): fetch AGENTS.md from base_sha, not head_sha
The reviewer inlines AGENTS.md into its system prompt. Reading from
head_sha means a PR author can edit AGENTS.md in the same PR being
reviewed and smuggle instructions like "ignore all bugs" / "publish
no findings" into the reviewer's prompt. Switch to base_sha (the
target branch's pre-PR state, which is trusted) and update the prompt
text to reflect that the contents come from the base, not the head.
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
The base reviewer prompt had grown into a curated catalog of defect
patterns and grep recipes tailored to the Martian eval set (Optional.get,
forEach(async), React keys, ERB templates, ORM increment, X-Frame-Options,
etc.). That bias toward one dataset hurts generalization across the
repos the reviewer actually runs on. Trim it down to universal review
discipline (the bar, do-not-file rules, severity rubric, publish
discipline, a generic workflow) and rely on per-repo prompts in memory
to carry repo-specific patterns at runtime.
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: TTL and revocation handling for cached GitHub OAuth tokens [closes AB-2322]
Persist github_token_expires_at alongside github_token_encrypted, treat
expired cache entries as missing so we re-resolve before kicking off
runs, and invalidate the cached ciphertext on a downstream 401 so the
next invocation gets a fresh token instead of replaying a revoked one.
Co-authored-by: Johannes du Plessis <51395795+johannes117@users.noreply.github.com>
* webapp: forward installation-token expiry to reviewer cache writes
The three reviewer-thread persist sites in webapp.py were calling
get_github_app_installation_token() (no expiry) and persist_encrypted_github_token
without expires_at, so cached App tokens were treated as never-expiring even
though they actually expire in ~1 hour.
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Johannes du Plessis <51395795+johannes117@users.noreply.github.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
