Stopgap for Vercel preview deploys rejecting open-swe[bot] PRs: even with
git config --global applied at sandbox setup, commits are landing with
open-swe <open-swe@langchain.dev> — likely from local repo config or shell
session env that overrides our global config.
Adding git config inside the cloned repo (local config) beats global and
ensures every commit picks up the bot's noreply email regardless of what
else is set in the sandbox.
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* fix: re-apply git identity on every agent invocation, source from env
Cached/reconnected sandboxes weren't getting `git config --global` re-applied
(the call was gated on new-sandbox creation), so commits picked up whatever
identity the sandbox happened to have — sometimes an email not associated
with any GitHub account, which Vercel rejects on preview deploys.
Move the git config call out of the new-sandbox branch so it runs on every
get_agent, and source the values from OPEN_SWE_GIT_AUTHOR_NAME /
OPEN_SWE_GIT_AUTHOR_EMAIL env vars (defaulting to the existing bot identity)
so forks can override without code changes.
* cleanup
---------
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
* feat: move github workflows to gh cli
Use LangSmith proxy auth to support gh-driven GitHub workflows while removing custom GitHub wrapper tools.
* docker ignore + snapshot and docker image updates
* updated image and instructions
* removing open_pr if needed after agent call
* feat: add edit_pull_request tool for editing PR titles and descriptions
* fix: patch auth flow in open PR middleware tests
* fix: support app token for editing PRs
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* feat: add release note section to PR body template
The langchainplus repo PR template includes a ## Release Note section,
but the agent hardcoded PR body template only had ## Description and
## Test Plan. This caused the release note section to be dropped from
agent-created PRs.
* docs: align PR body argument docs
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
When setting up the OAuth Provider in LangSmith using `github-oauth-provider` results in the error:
> Error: Provider ID 'github-oauth-provider' is reserved for Fleet public OAuth. Choose another name.
* fix: prevent futile retry loop when commit_and_open_pr fails with git/API errors
- Root cause: when git checkout or GitHub PR API fails, the tool returned a generic {"success": false} error with no signal that retrying is futile, causing the agent to loop 9-13+ times until hitting the 1000-step recursion limit
- Change: (1) git_checkout_branch now returns (bool, str) so the actual git error output is surfaced in the tool response; (2) checkout and PR creation failures now include "fatal": true and an explicit "Do not retry" message; (3) prompt.py COMMIT_PR_SECTION adds an explicit instruction to stop on fatal errors
- Verified: 109 unit tests pass, no regressions
* fix: skip PR safety net on fatal commit failures
* style(open_pr): ruff-format fatal retry skip condition
---------
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
- Root cause: LLM occasionally generates strings like '1, 80' or '170, "limit": 60'
for integer fields, causing a Pydantic ValidationError and wasting an LLM turn
- Change: add SanitizeToolInputsMiddleware in agent/middleware/sanitize_tool_inputs.py
that extracts the leading integer from any string value in offset/limit before
the call reaches Pydantic validation; registered before ToolErrorMiddleware in server.py
- Verified: 14 unit tests covering all three production trace patterns pass
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* fix: notify users via Slack when agent hits model call step limit
- Root cause: GraphRecursionError at 1000 steps bypassed all @after_agent
middleware including open_pr_if_needed, leaving users with no notification
- Change: Added ModelCallLimitMiddleware(run_limit=60) to intercept gracefully
before the hard recursion limit, and added notify_step_limit_reached
@after_agent middleware to post a Slack thread reply when the limit fires
- Verified: 107 existing tests pass, no regressions
* fix: harden step-limit Slack notification
Ensure the step-limit notification runs after the PR safety net and cover the new middleware behavior with focused unit tests.
---------
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* fix: warn against using http_request for GitHub PR creation
- Root cause: http_request tool description and system prompt lacked explicit guidance not to use it for GitHub PR operations, causing the agent to fall back to it and receive 401 Unauthorized responses
- Change: added clear warnings to both the http_request docstring and the TOOL_USAGE_SECTION in prompt.py directing agents to use commit_and_open_pr instead
- Verified: docstring and prompt changes are minimal and scoped
* fix: clarify http_request PR guidance
---------
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* fix: prevent scope creep — clarify lockfile handling and task scope guidance
- Root cause: CODING_STANDARDS_SECTION ambiguously said 'package manager files' (interpreted to include lockfiles like uv.lock, package-lock.json) and TASK_EXECUTION_SECTION scope guidance was too generic to prevent cross-language or cross-service modifications
- Change: clarify that only manifest files (pyproject.toml, package.json) need updating and auto-generated lockfiles must NOT be committed; add concrete examples to scope guidance (no JS when Python-only, no other services when one is targeted)
- Verified: 3 production traces showing users complaining about scope creep
* fix: clarify prompt lockfile guidance
---------
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
- Root cause: create_github_pr found an existing PR on 422 but never
PATCHed it, so callers like commit_and_open_pr could not update the
PR body (e.g. adding "Closes AB-1159") on subsequent invocations.
- Change: after _find_existing_pr succeeds, call new _update_github_pr
helper which PATCHes /repos/{owner}/{repo}/pulls/{number} with the
requested title and body before returning pr_existing=True.
- Verified: self-evident API call addition; proof in production traces.
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* fix: return actionable error when git push fails due to workflows scope
- Root cause: push failure with "workflows scope" error returned a generic
"Git push failed: ..." message, causing the agent to retry 10+ times
- Change: detect "workflows" + "scope" in push output and return a clear
message instructing the agent to remove .github/workflows/ file changes
- Verified: unit tests cover both the workflow-scope path and the
non-workflow path to prevent regressions
* fix: detect github workflow permission push failures
* style: ruff-format checkout line in commit_and_open_pr
---------
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
- Root cause: httpx.HTTPError handler returned (None, None, False) without
checking if the PR was already created on GitHub before the network error
- Change: added _find_existing_pr fallback in except block in agent/utils/github.py
- Verified: 5 production traces showed false failures where PR existed (pr_existing=True on retry)
Co-authored-by: LangSmith Forge <forge-agent@langsmith.ai>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* fix: stop agent retrying commit_and_open_pr on 403 permission denied
Detect 403/permission-denied push failures in commit_and_open_pr and
return a PERMANENT_FAILURE message so the LLM stops retrying. Also add
prompt-level guidance to the COMMIT_PR_SECTION reinforcing this. Add
unit tests covering both the 403 and non-403 push failure paths.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: stop safety net retrying permanent push failures
Skip the after-agent PR fallback when commit_and_open_pr reports a permanent GitHub push authorization failure, while preserving fallback behavior for recoverable failures.
---------
Co-authored-by: Claude Agent <agent@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* feat: add github CI check run tools for shepherding CI
Add get_pr_check_runs and rerun_failed_check_runs tools that authenticate
using the GitHub App installation token so the agent can query and retry
CI status on private repos without relying on GH_TOKEN or unauthenticated
http_request calls.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: handle paginated GitHub CI results
* refactor(github_ci): address review feedback
- Rename rerun_failed_check_runs -> rerun_failed_workflow_runs and clarify
in docstrings that the tool only retries GitHub Actions workflow runs
(not third-party CI checks surfaced by get_pr_check_runs).
- Skip action_required workflow runs when rerunning; those need manual
approval, not a rerun.
- Run rerun-failed-jobs requests concurrently via asyncio.gather instead
of sequentially.
- Fix latent pagination bug in _fetch_paginated_items where caller-supplied
params could overwrite per_page/page and break the end-of-pagination
check; reserved keys now always win and the threshold uses a PER_PAGE
constant.
- Set an explicit 30s httpx timeout so a hung GitHub call cannot stall
the agent loop.
- Restore alphabetical ordering of tools in agent/tools/__init__.py.
- Add tests for: a 500 surfaced on a later pagination page, and
action_required runs being filtered out of rerun candidates.
---------
Co-authored-by: Claude Agent <agent@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
When the agent is triggered by automated deployment monitoring (no Slack
thread, no Linear ticket, no GitHub issue), the prompt now explicitly
instructs it to skip the notification step rather than attempting
slack_thread_reply and failing.
Co-authored-by: Claude Agent <agent@anthropic.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Calling get_github_token() without arguments always invoked LangGraph get_config internally,
which broke tests that only patch agent.middleware.open_pr.get_config and failed outside
runnable context.
Extend get_github_token with an optional runnable config mapping; the middleware passes
the config dict already resolved from get_config(). Request GitHub App installation tokens
only after detecting sandbox/repo changes worth publishing.
Fixes failing Agent unit tests in tests/test_open_pr_middleware.py.
* fix: safety net middleware always skipped due to key-existence check
The open_pr_if_needed after-agent middleware checked `if 'success' in pr_payload`
which evaluates True for BOTH success and failure responses from commit_and_open_pr
(all responses include the 'success' key). This meant the safety net never fired.
Fix: use `pr_payload.get('success')` to check the VALUE instead of key existence.
Evidence: 6+ production traces in last 24h where commit_and_open_pr returned
success=False but the safety net silently skipped (non-fast-forward push failures,
missing GitHub token, workflow permission errors, API 500 errors).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* update
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Palash Shah <palash@langchain.dev>
* feat: add get_pr_review_comments tool for authenticated GitHub API access
The agent was asking users to paste PR review comments because it had no
tool to fetch them with auth. This adds get_pr_review_comments, which uses
the GitHub App installation token to fetch all three comment types (thread
comments, inline review comments, review submissions) from private repos.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* u
* u
---------
Co-authored-by: Forge Agent <agent@forge.ai>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Palash Shah <palash@langchain.dev>
Co-authored-by: Palash Shah <35114859+Palashio@users.noreply.github.com>
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Previous defaults (4 vCPU / 15 GiB) exceed the maximum sandbox size
the LangSmith API accepts, causing a 400 on every sandbox creation:
sandbox size 4 vCPU / 15360 MiB exceeds the maximum supported size;
must fit within one of: small (1 vCPU / 1792 MiB),
medium (1 vCPU / 3840 MiB), or large (2 vCPU / 7936 MiB)
Default to the "large" cap (2 vCPU / 7936 MiB) so deployments without
DEFAULT_SANDBOX_VCPUS / DEFAULT_SANDBOX_MEM_BYTES env overrides boot
into a working state.
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: open PRs under user's name and add OpenSWE label
* feat: use user token for PR authorship, add OpenSWE label, and consolidate fallback logic
* linting
* fix: address review nits for PR authorship and labeling
Fix docstring casing, add debug logging for 422 existing-PR search
fallback, tighten test type annotations, and add missing HTTPError
fallback test.
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: default to GPT-5.5 medium reasoning
Use OpenAI GPT-5.5 with medium reasoning as the default model and document the completion-token budget semantics for reasoning models.
* fix: use Responses API reasoning config
Pass GPT-5.5 reasoning settings through LangChain's Responses API parameter instead of the Chat Completions-only reasoning_effort field.
* feat: raise GPT-5.5 output budget
Set the default GPT-5.5 output token budget to the model maximum so long-running coding tasks have more room for reasoning and final responses.
* feat: align recursion limit with Deep Agents
Use Deep Agents' default recursion limit so longer coding runs have room to complete without Open SWE imposing a lower cap.
* chore: remove minimal effort level
* chore: reduce max tokens to 64_000
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
* feat: migrate LangSmith sandbox creation to snapshot API
Replaces the template-based sandbox flow (DEFAULT_SANDBOX_TEMPLATE_NAME /
DEFAULT_SANDBOX_TEMPLATE_IMAGE) with the new snapshot-based flow.
- New required env var DEFAULT_SANDBOX_SNAPSHOT_ID (UUID of a pre-built
LangSmith snapshot; build out-of-band via UI or SandboxClient.create_snapshot)
- Optional DEFAULT_SANDBOX_SNAPSHOT_FS_CAPACITY_BYTES overrides the root FS
size at boot (default 32 GiB)
- Startup-time validation via a FastAPI lifespan hook: the server refuses
to boot with a clear ValueError if SANDBOX_TYPE=langsmith and
DEFAULT_SANDBOX_SNAPSHOT_ID is unset, so failures surface in boot logs
rather than on the first thread
- Reconnect-to-existing-sandbox path unchanged
- Docs (INSTALLATION.md, CUSTOMIZATION.md) updated to describe the new
snapshot workflow
* fix: format create_sandbox_snapshot.py to pass ruff
---------
Co-authored-by: aran-yogesh <yogesh.mahendran@langchain.dev>
