* 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>
Additive sandbox provider adapted to this fork's synchronous create_sandbox
factory: agent/integrations/e2b.py registered as a lazy-import entry in
SANDBOX_FACTORIES, so the e2b/langchain-e2b imports only load when
SANDBOX_TYPE=e2b (dark-safe on unset env; unset E2B_API_KEY raises a clean
ValueError only when the provider is selected). Supports reconnect-by-id and
optional E2B_TEMPLATE. Non-langsmith providers skip the GitHub proxy step,
so the GitHub-App token flow is untouched.
Upstream: langchain-ai/open-swe 48217b68 (#1489), re-implemented against the
fork's sync sandbox lifecycle rather than cherry-picked (upstream ships on
the deferred async sandbox.py base).
Docs: provider tables/lists in README, CUSTOMIZATION, INSTALLATION; the
CUSTOMIZATION registration example now shows the lazy-tuple form the code
actually uses. uv.lock refreshed (adds e2b, langchain-e2b, dockerfile-parse).
* fix: migrate stale admin model defaults
Normalize retired model IDs before validation and in settings responses so full admin updates remain saveable after model upgrades.
* fix: restrict retired model migration
Only migrate explicitly retired model IDs so malformed provider model names and efforts continue to fail validation.
(cherry picked from commit 62e0ca2d4898ebc3c2ae8887030a364779d907cb)
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
(cherry picked from commit 092abafa4cb955c3823f727d35d7bad94e1147ab)
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* fix: let admins interrupt runaway agents (#1730)
Add a workspace-wide admin control that cancels every active run while preserving thread history.
(cherry picked from commit 09eaf94c3e612db969daa000b909805963fe1de9)
* chore: reconcile triage ledger
Signed-off-by: Adam Moussa <adam@seahavenind.com>
* fix: change ui import from refactor
Signed-off-by: Adam Moussa <adam@seahavenind.com>
---------
Signed-off-by: Adam Moussa <adam@seahavenind.com>
Co-authored-by: Johannes du Plessis <johannes@langchain.dev>
* feat(open-swe): add Jira tool plane (Phase 1)
Curated Jira Cloud REST v3 toolset for the agent, mirroring the Linear
tools:
- utils/jira.py: service-account REST client (Basic auth) with get/
create/update issue, comments, list projects, trace comment; issue and
comment bodies normalized to markdown.
- utils/adf.py: minimal ADF <-> markdown conversion (read paths convert
Jira ADF to markdown; agent comments convert prose to ADF).
- tools/jira_{comment,get_issue,get_issue_comments,create_issue,
update_issue,list_projects}.py wired into the tool registry and the
main agent tool list.
- tests/test_jira_utils.py: ADF conversion + mocked-transport util tests.
Reads JIRA_BASE_URL / JIRA_SERVICE_EMAIL / JIRA_API_TOKEN; unset env
returns a clean error, so this is safe to land dark. Trigger plane,
prompt guidance, and config plumbing follow in Phase 2.
* feat(open-swe): add Confluence tool plane (Phase 3)
Curated Confluence Cloud REST toolset for the agent, mirroring the Jira
tools:
- utils/confluence.py: service-account REST client (Basic auth) with
get/create/update page, add comment, CQL search. Page bodies are XHTML
storage format (not ADF), with minimal storage<->text converters;
update_page reads the current version and bumps it, as Confluence
requires.
- tools/confluence_{get_page,create_page,update_page,comment,search}.py
registered in the tool registry.
- tests/test_confluence_utils.py: converter + mocked-transport tests
including the version-bump path.
Reads CONFLUENCE_BASE_URL / CONFLUENCE_EMAIL / CONFLUENCE_API_TOKEN;
unset env returns a clean error. Activation in the agent tool list lands
with the Phase 2 server.py wiring.
* feat(open-swe): add Jira trigger plane (Phase 2)
Make an @openswe comment on a Jira issue spawn an agent run, mirroring
the Linear trigger plane:
- webhooks/jira.py: process_jira_issue clones process_linear_issue —
deterministic thread id, full-issue fetch, actor accountId->email
attribution feeding resolve_login_from_email_async (PRs open as the
human), multimodal image handling, source="jira" + jira_issue config.
- webapp.py: POST/GET /webhooks/jira, verify_jira_secret (constant-time
X-Automation-Webhook-Token check, fails closed), repo-resolution
cascade, get_repo_config_from_jira_mapping.
- utils/jira_project_repo_map.py: JIRA_PROJECT_TO_REPO (placeholder
entry — real project->repo mappings still needed).
- utils/jira.py: get_user_email (accountId -> email) for attribution.
- completion.py: source=="jira" failure-reply branch.
- prompt.py: Jira-triggered notify guidance + Refs:/branch key from
{jira_project_key}-{jira_issue_number}.
- server.py: read jira_issue config + pass jira key to the system
prompt; also activates the Phase 3 Confluence tools in the agent list.
Jira Automation lacks native webhook HMAC signing, so trust is a shared
secret header (decision D2); replay protection is weaker than Linear's
HMAC+timestamp. /sh-security-review + an Atlassian IP allowlist are the
outstanding gate/hardening before push.
* fix(open-swe): harden Jira webhook trust (sh-security-review)
Resolves findings from the Phase 2 security review (detector fan-out +
proof-or-kill verifier). The unsigned Jira Automation webhook body was
trusted for identity, comment content, repo routing, and issue
existence; a JIRA_WEBHOOK_SECRET holder could forge those fields.
- Corroborate against the real Jira record: the webhook body is now only
a pointer (issue_key + required comment_id). The triggering comment's
author and text are re-fetched server-side via get_comment/fetch_jira_
comment, and identity, the @openswe check, prompt text, and project
key are derived from that authoritative record — never payload author/
body fields. An uncorroborated comment is rejected. (closes the
account-id impersonation, unsigned-body prompt injection, and
fabricated-issue findings)
- Validate issue_key against the Jira key format and percent-encode all
untrusted path segments (_seg) so a crafted key can't traverse to a
different Jira REST endpoint or inject query params. (closes the path-
traversal / query-injection findings)
- Route source=="jira" through the bot-token-default / author_prs_as_
user opt-in path in resolve_github_token, matching Linear, instead of
unconditionally resolving a per-user OAuth token from a payload email.
- Gate attribution on an active user mapping (is_login_mapped) so a
pending/unconfirmed mapping can't drive PR authorship.
Adds regression tests: server-corroboration wins over payload, malformed
issue_key rejected, uncorroborated comment rejected, path-segment
encoding, project-key derivation, active-mapping gate.
Remaining (non-blocking, deployment/hardening): set ALLOWED_GITHUB_ORGS/
REPOS so the shared allowlist isn't fail-open; consider HMAC-over-body +
timestamp on the Automation payload to close the residual replay gap.
* harden(open-swe): opt-in Jira webhook replay/IP + fail-closed allowlist
Folds the two deployment-hardening items from the Phase 2 security review
into code (all opt-in / default-off, so existing and upstream deployments
are unaffected):
- JIRA_WEBHOOK_REQUIRE_SIGNATURE: when set, the Automation payload must
carry X-Openswe-Signature (hex HMAC-SHA256 of the raw body keyed by
JIRA_WEBHOOK_SECRET) plus a fresh timestamp, verified by
verify_jira_signature / _jira_timestamp_is_fresh (mirrors the Linear
HMAC+freshness model). Closes the static-token model's replay/forgery
gap when enabled.
- JIRA_WEBHOOK_IP_ALLOWLIST: optional CIDR allowlist on the webhook's
direct client IP (verify_jira_source_ip). Documented as direct-peer
only; behind a proxy/LB, allowlist Atlassian's ranges at that layer.
- REQUIRE_REPO_ALLOWLIST: makes an empty ALLOWED_GITHUB_ORGS/REPOS fail
CLOSED instead of the back-compat allow-all, plus a startup fail-open
warning. Applies to all channels for consistency.
Documents all new vars (and a Jira section) in .env.example. Adds tests
for signature on/off + valid/missing/wrong/stale, IP allow/deny/off, and
the fail-closed allowlist.
* feat(open-swe): Confluence Atlassian Connect trigger (Phase 4)
Adds the @openswe-on-a-Confluence-comment trigger via a private Atlassian
Connect app. Designed and adversarially verified with the ultracode
workflow (3 divergent Opus designs + judge; 3 proof-or-kill Opus
skeptics on the implemented crypto).
- utils/atlassian_connect.py: hand-rolled qsh (pinned to Atlassian's
official test vector), PyJWT HS256 webhook verifier with alg-pinning,
issuer binding, and qsh-verified-last ordering; RS256 signed-install
lifecycle verifier against Atlassian's published keys; installation
store keyed by clientKey with the sharedSecret encrypted at rest
(TOKEN_ENCRYPTION_KEY / Fernet). No new dependency (PyJWT already pinned).
- webhooks/confluence.py: install/uninstall lifecycle + comment handler.
The JWT-signed webhook body is only a pointer; the comment's real
author/text/container are re-fetched server-side via the Basic-auth
service account (Phase-2 corroboration lesson), with active-only login
attribution and the repo allowlist.
- utils/confluence.py: get_comment / get_user_email (path-encoded).
- webapp.py: GET /connect/atlassian-connect.json (served dynamically),
POST /connect/{installed,uninstalled,webhook/comment-created}, the
space->repo resolver, thread-id, and fetch helpers.
- completion.py: source=="confluence" failure-reply branch.
Security: the sh-security-review verify pass confirmed one HIGH — the
symmetric signed-install=false first-install was trust-on-first-use gated
only by the public Confluence hostname (webhook-auth bypass). Fixed by
switching to signed-install=true + RS256 verification of lifecycle
callbacks, which cryptographically authenticates the first install. All
other attack lenses (forgery/replay/alg-confusion/overwrite/uninstall
DoS/corroboration/injection) were defeated; residuals are deployment
config (REQUIRE_REPO_ALLOWLIST) or accepted-by-design (qsh cannot cover
bodies; comment-trigger prompt injection, shared with all sources).
New env (documented in .env.example): CONFLUENCE_BASE_URL/EMAIL/API_TOKEN,
CONNECT_BASE_URL, CONNECT_EXPECTED_BASE_URL (optional). Install secrets
require the durable Postgres LangGraph store in prod.
Outstanding before push: /sh-security-review on the real diff and the
GPT-4.1 cross-family review (auth boundary); README/CLAUDE.md + memory.
* docs(open-swe): Phase 5 — Confluence prompt guidance + architecture docs
- prompt.py: Confluence-triggered runs notify via confluence_comment on
the triggering page; add Confluence to the shared-base source list.
- CLAUDE.md: document the Jira + Confluence tool planes and the Atlassian
triggers (Jira Automation shared-secret webhook; Confluence Connect app
with HS256 webhook + qsh and RS256 signed-install lifecycle), plus the
server-side corroboration + encrypted install store.
Phase 5 also verified the trigger surface end-to-end against a running
uvicorn app (descriptor served; /connect/* and /webhooks/jira fail closed
without valid auth) and recorded the integration in project memory.
* fix(open-swe): resolve /sh-security-review findings on the Atlassian surface
Formal sh-security-review (detector fan-out + verifier) over the Phase-4
Connect surface (esp. the new RS256 signed-install code, unseen by the
earlier adversarial verify) and the Phase-2 opt-in hardening.
CRITICAL — cross-tenant install (origin validation, CWE-346): signed-
install proves the caller is *an* Atlassian tenant, not *ours*, and the
descriptor is served publicly, so any attacker could install the app on
their own Confluence site and drive agent runs against our allowlisted
repos. The baseUrl body field is attacker-controlled and cannot bind the
tenant; only the signature-verified clientKey (JWT iss) can. Added a
MANDATORY, fail-closed CONNECT_EXPECTED_CLIENT_KEYS allowlist checked in
process_install after signature+iss verification.
HIGH — cross-tenant thread-id collision (CWE-330/863): Confluence comment
ids are per-instance, so generate_thread_id_from_confluence_comment now
salts the hash with the verified clientKey (plumbed from the webhook JWT
iss) to prevent thread hijack across tenants.
HIGH/MEDIUM — path/query injection (CWE-22/88): get_page and update_page
interpolated page_id into the REST path unencoded (update_page on a
mutating PUT with no params= backstop). Now _seg()-encoded, matching the
rest of the module.
MEDIUM — self-trigger loop (CWE-405): process_confluence_comment had no
bot-authorship early-out. Added an optional CONFLUENCE_BOT_ACCOUNT_ID
guard mirroring the Linear botActor / Jira comment_author_is_bot checks.
LOW — corrected the CONNECT_EXPECTED_BASE_URL comment to document it as
opt-in defense-in-depth (the clientKey allowlist is the real gate).
Verified clean by the detectors: RS256/HS256 alg-pinning, aud/iss/exp,
kid-fetch SSRF (host-pinned + quote-encoded), at-rest secret encryption,
constant-time comparisons, and the Phase-2 hardening. New regression
tests for each fix; full suite green (1602).
* harden(open-swe): GPT-4.1 cross-family review follow-ups
Cross-family review (GPT-4.1 via orchestrator cross_reviewer) found no
critical/high issues and confirmed the auth boundary is fail-closed and
correct. Two low-cost defense-in-depth items applied:
- Validate the signed-install JWT 'kid' against a strict charset before
the public-key fetch, so a malformed kid fails fast with no network
call (on top of the existing fixed host + percent-encoding).
- Make JWT nbf verification explicit (verify_nbf) on both the RS256
lifecycle and HS256 webhook decodes.
Other suggestions triaged as already-handled (aud cross-app replay is
blocked by the per-tenant iss->secret lookup; documented static-token/IP/
baseUrl tradeoffs; qsh pinned to Atlassian's vector) or ops/infra
(Fernet rotation via MultiFernet; rate limiting at the gateway).
* docs(open-swe): document Jira + Confluence in installation & customization guides
- INSTALLATION.md §5: add Jira (Automation-rule webhook + shared secret,
service account, JIRA_PROJECT_TO_REPO) and Confluence (Atlassian Connect
app install, CONNECT_EXPECTED_CLIENT_KEYS bootstrap, durable-store note,
CONFLUENCE_SPACE_TO_REPO) trigger setup; §6: add the new env vars +
REQUIRE_REPO_ALLOWLIST.
- CUSTOMIZATION.md: jira_*/confluence_* in the tools table; repo-extraction
note covers all four sources.
- AGENTS.md: match CLAUDE.md (triggers, webhooks, tool list, auth).
- README.md: invocation section, tools table, and overview line.
* feat: open Linear-triggered PRs as the triggering user
Add 'linear' to the set of sources that carry a mapped GitHub login
(Slack, Linear, dashboard, schedule), so Linear-triggered runs resolve
the per-user OAuth token when the author_prs_as_user profile flag is
enabled. Previously only Slack and dashboard runs could open PRs as the
user; Linear runs always used the bot token.
The Linear webhook now resolves the GitHub login from the Linear email
via the same user-mapping store Slack uses, and passes it through the
run configurable and thread owner metadata.
Refs: 5003c953 (upstream #1683)
* fix(open-swe): restrict Linear token attribution to comment author only
Split actor_email (comment_author only, feeds github_login for token
attribution) from user_email (full fallback chain, for display/model).
This ensures a PR is never opened as a non-actor (creator/assignee).
Add test_resolve_github_token_linear_defaults_to_bot to lock the
security-critical default: Linear + mapped login + no opt-in → bot.
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
* feat: surface attributed PR creation failures
Port upstream #1659: adds PullRequestCreationGuardMiddleware that
blocks shell fallbacks (gh pr create, gh api /pulls, curl) when
open_pull_request fails, keeping failures visible. Also adds preflight
branch/repo visibility checks in open_pull_request with structured
failure payloads, and updates the prompt to forbid PR creation
fallbacks.
Refs: #134
* fix: fall back to core GitHub App scope when optional grants missing (#1701)
* fix: fall back to core GitHub App scope when optional grants missing
Proxy-token minting requested workflows:write and actions:read in the
permission set used for every sandbox. GitHub 422s a token request that
asks for a permission the installation hasn't granted, so any install
without workflows:write failed to mint a token and every run died in
before-agent setup with "GitHub App installation token is unavailable".
_resolve_proxy_token now walks a permission ladder (full -> +workflows ->
core) and returns the first scope that mints, recording the granted scope
so hourly proxy refreshes stay consistent. A missing optional grant now
degrades to the install-time core scope instead of failing the run;
workflow-file HITL pushes still require workflows:write and fail at push
time when it is absent.
* refactor: flatten proxy-token ladder loop with continue
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
(cherry picked from commit f53caff1aa24a7b29d851b267aa3bdfe62c1e935)
Sea Haven fork deviation: upstream #1701 folds workflows:write into the
standing BASE/RUNTIME scope. This fork deliberately keeps workflows:write
OUT of the standing permission ladder (RUNTIME = core + actions:read;
LADDER = (RUNTIME, CORE)) so the sandbox proxy token cannot push
.github/workflows/* during normal operation. workflows:write is minted only
transiently by WorkflowPushGuardMiddleware for an approved HITL push and
dropped on restore, preserving token scope as a backstop for the workflow-
push approval control. Security-reviewed (agentic fan-out + GPT-4.1 cross
review); the standing-scope-carries-workflows:write bypass was blocked.
* fix(open-swe): harden proxy-token restore and mint error handling
Two low-severity follow-ups from the security review of the #1701 port.
Restore the recorded baseline scope after a workflow-push elevation instead
of a hardcoded RUNTIME. An install granted workflows:write but not actions:read
resolves its standing token to core; hardcoding RUNTIME on restore requested the
ungranted actions:read, 422'd, and fired a false "SECURITY: failed to downscope"
error on every approved workflow push before the core fallback recovered. The
guard now captures the run's recorded scope before elevating (via the new
get_recorded_proxy_permissions) and restores exactly that, falling back to the
guaranteed core scope only when the baseline restore fails.
Classify installation-token mint failures. get_github_app_installation_token_
with_expiry now treats HTTP 422 (a permission the installation hasn't granted)
as the ladder's expected descend signal and keeps it at debug, while a non-422
failure (network/5xx/timeout) is surfaced at WARNING even when errors are
otherwise suppressed — so a transient blip no longer silently downscopes a whole
run under a debug-only trace. The reduced-scope warning no longer asserts a
missing grant as the sole cause.
* chore(triage): mark upstream #1701 landed on this branch
Ported via PR #181 as Option A (workflows:write kept out of the standing
proxy-token scope). Regenerated triage.md from triage.jsonl.
* fix: restructure PR creation to POST-first with diagnose-on-failure
Move preflight checks from an authoritative gate (before POST) to a
diagnostic run after POST failure. This avoids false-positive failures
when a just-pushed head branch is momentarily invisible to GitHub ref
endpoints, and eliminates 2-3 extra serial API round-trips on the happy
path.
Also drop unused _PR_CREATED_FALSE indirection and add a docstring to
pr_creation_guard acknowledging the fail-open detection design.
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
Co-authored-by: Ramon Nogueira <ramon.nogueira@langchain.dev>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
* fix: fall back to core GitHub App scope when optional grants missing (#1701)
* fix: fall back to core GitHub App scope when optional grants missing
Proxy-token minting requested workflows:write and actions:read in the
permission set used for every sandbox. GitHub 422s a token request that
asks for a permission the installation hasn't granted, so any install
without workflows:write failed to mint a token and every run died in
before-agent setup with "GitHub App installation token is unavailable".
_resolve_proxy_token now walks a permission ladder (full -> +workflows ->
core) and returns the first scope that mints, recording the granted scope
so hourly proxy refreshes stay consistent. A missing optional grant now
degrades to the install-time core scope instead of failing the run;
workflow-file HITL pushes still require workflows:write and fail at push
time when it is absent.
* refactor: flatten proxy-token ladder loop with continue
---------
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
(cherry picked from commit f53caff1aa24a7b29d851b267aa3bdfe62c1e935)
Sea Haven fork deviation: upstream #1701 folds workflows:write into the
standing BASE/RUNTIME scope. This fork deliberately keeps workflows:write
OUT of the standing permission ladder (RUNTIME = core + actions:read;
LADDER = (RUNTIME, CORE)) so the sandbox proxy token cannot push
.github/workflows/* during normal operation. workflows:write is minted only
transiently by WorkflowPushGuardMiddleware for an approved HITL push and
dropped on restore, preserving token scope as a backstop for the workflow-
push approval control. Security-reviewed (agentic fan-out + GPT-4.1 cross
review); the standing-scope-carries-workflows:write bypass was blocked.
* fix(open-swe): harden proxy-token restore and mint error handling
Two low-severity follow-ups from the security review of the #1701 port.
Restore the recorded baseline scope after a workflow-push elevation instead
of a hardcoded RUNTIME. An install granted workflows:write but not actions:read
resolves its standing token to core; hardcoding RUNTIME on restore requested the
ungranted actions:read, 422'd, and fired a false "SECURITY: failed to downscope"
error on every approved workflow push before the core fallback recovered. The
guard now captures the run's recorded scope before elevating (via the new
get_recorded_proxy_permissions) and restores exactly that, falling back to the
guaranteed core scope only when the baseline restore fails.
Classify installation-token mint failures. get_github_app_installation_token_
with_expiry now treats HTTP 422 (a permission the installation hasn't granted)
as the ladder's expected descend signal and keeps it at debug, while a non-422
failure (network/5xx/timeout) is surfaced at WARNING even when errors are
otherwise suppressed — so a transient blip no longer silently downscopes a whole
run under a debug-only trace. The reduced-scope warning no longer asserts a
missing grant as the sole cause.
* chore(triage): mark upstream #1701 landed on this branch
Ported via PR #181 as Option A (workflows:write kept out of the standing
proxy-token scope). Regenerated triage.md from triage.jsonl.
---------
Co-authored-by: Ramon Nogueira <ramon.nogueira@langchain.dev>
Port upstream 304032fa: clarify that information-only answers should
check out relevant repos first for context, answer fully inline, and
only post a concise summary to Slack threads.
The fork already carried the shared-base Slack guidance; this adds the
missing TASK_EXECUTION_SECTION update and its test coverage.
Refs: #146
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
The WorkflowPushGuardMiddleware fails closed and blocks a git push whenever
it cannot parse the command into a plain, inspectable form (chained shell
operators, obfuscation, unrecognized refspecs). Most block reasons named only
the symptom ("chained shell commands"), so the agent kept retrying other
chained variants (cd &&, pushd &&) instead of dropping the chaining.
Append a single actionable remedy to every block reason, pointing at a plain
`git push origin <branch>` or `git -C <dir> push`, so a blocked run recovers
on the next attempt instead of looping.
* feat(models): re-add Fable 5 with admin disable toggle (port of upstream #1677)
* refactor(models): convert re-added Fable 5 to Bedrock model IDs
* fix(open-swe): correct Fable copy to describe provider data sharing, not ZDR
The ported admin toggle description and code comments described Fable 5 as
incompatible with Zero Data Retention. That is backwards: Fable 5 requires
the account to opt into Bedrock provider_data_share — prompts/completions are
retained and shared with Anthropic (up to 30 days, incl. human review). The
old UI copy would lead an admin to believe the opposite of what enabling the
toggle does. Reword the toggle description and the gate_fable_model /
team_settings comments accordingly. Still off by default. Refs #171.
* feat(dashboard): surface thread sandbox ID in sidebar
Port of upstream #1689 (feb7ac98): expose the thread's sandbox_id on the
dashboard thread summary and surface it in the sidebar via a "Copy sandbox
ID" action.
- thread_api: add sandboxId to the thread summary, hiding the
"__creating__" in-flight sentinel.
- queries/types: thread the sandboxId field through AgentThread and the
optimistic thread.
- AgentsSidebar: replace the hover-only resolve/delete buttons and the
right-click context menu with one touch-friendly kebab (⋮) menu that
works on both pointer and touch, and add the Copy sandbox ID item.
- vite: disable the PWA service worker in dev (it precaches assets and
defeats HMR).
- tests: unit test for the summary field + e2e spec in the real-backend
harness.
Closes#140. Part of #134.
* chore(triage): mark feb7ac98 (#1689) landed
Ported to dev via feat/thread-sandbox-id-sidebar (#140).
dev already pins fireworks-ai>=1.2.0a88 (uv.lock resolves a88) via #158
(efeb6b12), which jumped a85 -> a88 and exceeds the a86 that upstream
#1667 (fbc6de85) bumps to. Porting it would downgrade the constraint, so
mark the row landed/superseded rather than picking it.
Closes#141. Part of #134.
The built-in GITHUB_TOKEN cannot open the sync PR: the enterprise policy
blocks GitHub Actions from creating/approving pull requests, which
overrides the org and repo settings. That restriction applies only to
github-actions[bot], so mint a PROMOTE_APP installation token and pass
it to create-pull-request, mirroring promote-to-main.yml.
Add upstream-ledger-sync action to run daily to pull upstream commits, update triage.jsonl, and open a PR against dev so the list can be automatically updated to stay in line with upstream.
* feat: port plan-review & workflow-approval UX (#135)
Port six upstream commits onto dev:
- c03a6be7 (already ported): keep plan guidance high-level
- 546042a4: add workflow approval UI with diff preview, approval URLs,
web review links, and polling for approval status during active runs
- 216cf181: remove workflow token elevation; approved pushes pass
through directly without proxy token rewriting
- 3dbc0282: preserve plan redirects after login by accepting relative
same-origin redirect_to values and rejecting blocked paths
- bb104d93: submit plan comments with cmd+enter
- 90cb6caa: terse Slack replies, shared content via save_plan outside
plan mode (PLAN_STATUS_SHARED), reject shared-content mutations
Refs: #135
* feat: port durable dispatch hardening and startup latency improvements
Port five upstream PRs onto dev:
- #1621 / #1658: durable dispatch with loopback webhook defense,
create_durable_run helper, _config_with_prepare_run_id, degradation
to None for relative/loopback completion webhook URLs
- #1696: run-level completion webhook deduplication (replace
claim-then-post with post-then-flag per run_id), DeferredErrorModel
for graph-factory resilience, ToolRetryMiddleware for task subagents,
TimeoutWrapupMiddleware for all three graphs
- #1697: lazy-load __init__.py for agent.middleware, agent.tools,
agent.dashboard (PEP 562); defer heavy imports (exa_py in web_search,
agent.webapp in request_pr_review, deepagents in sandbox.py); add
ttl_cache.py with stale-while-revalidate for tool loaders
Refs: #137
* fix: restore login page render and clear CI lint/format
The plan-review port removed the authRedirectUrl import from login.tsx
but left its call site, crashing the login page at runtime (blank page,
no 'Sign in to open-swe'). Pass the relative path straight to loginUrl,
matching the plan route and the backend relative-redirect handling.
Also drop an unused os import in the guard test and reformat
workflow_push_guard.py to satisfy ruff.
* feat: port model-fallback resilience from upstream (#1694, #1695)
- Add httpx.TransportError to the transient-exception set so
incomplete chunked reads on streamed responses trigger a
fallback instead of cancelling the run (#1694 / c9f6dd86).
- Rewrite fallback to alternate primary/fallback with exponential
backoff instead of a single failover, so the agent survives
multi-minute gateway outages spanning both providers (#1695 /
c9a9a7cd).
- Default backoff schedule (0, 5, 15, 30, 45) reaches past the
gateway's ~30s recovery window; jittered ±25%.
- On exhaustion, surface a terminal AIMessage explaining the
outage instead of crashing — progress is checkpointed so the
user can retrigger to continue.
- Preserve fork conventions: Bedrock ClientError retryability
check, sync wrap_model_call (using time.sleep instead of
asyncio.sleep), and existing access-error surfacing for
Anthropic, OpenAI, and Bedrock (botocore) provider errors.
- Update triage ledger (c9f6dd86, c9a9a7cd → landed) and
re-render triage.md.
Refs: #139
* fix: align workflow-push-guard tests with dev's transient-elevation impl
The dev merge auto-combined dev's elevation tests with the stale passthrough
tests inherited from the durable-dispatch branch; the passthrough tests
contradict dev's restored _run_with_workflow_token impl. Take dev's test file.
* fix: drop dead ttl_cache module; make fallback backoff jitter two-sided
ttl_cache.py was re-introduced via the dev merge but dev/#160 deliberately
removed it as dead code (no agent importer). Remove it to match dev.
Also make _jittered_delay symmetric (±25%) to match its docstring.
* chore(upstream-sync): triage 4 new upstream commits (#1708-#1713)
Synced ledger to upstream/main (71e3b818). New rows all deferred:
- #1708 add GPT-5.6 OpenAI models (FLAG-HUMAN: fork picker is Bedrock/Fireworks-only)
- #1709 stale admin model defaults after upgrades
- #1710 bump langchain-fireworks 1.4.4
- #1713 align reviewer eval with published findings
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
* feat: port plan-review & workflow-approval UX (#135)
Port six upstream commits onto dev:
- c03a6be7 (already ported): keep plan guidance high-level
- 546042a4: add workflow approval UI with diff preview, approval URLs,
web review links, and polling for approval status during active runs
- 216cf181: remove workflow token elevation; approved pushes pass
through directly without proxy token rewriting
- 3dbc0282: preserve plan redirects after login by accepting relative
same-origin redirect_to values and rejecting blocked paths
- bb104d93: submit plan comments with cmd+enter
- 90cb6caa: terse Slack replies, shared content via save_plan outside
plan mode (PLAN_STATUS_SHARED), reject shared-content mutations
Refs: #135
* feat: port durable dispatch hardening and startup latency improvements
Port five upstream PRs onto dev:
- #1621 / #1658: durable dispatch with loopback webhook defense,
create_durable_run helper, _config_with_prepare_run_id, degradation
to None for relative/loopback completion webhook URLs
- #1696: run-level completion webhook deduplication (replace
claim-then-post with post-then-flag per run_id), DeferredErrorModel
for graph-factory resilience, ToolRetryMiddleware for task subagents,
TimeoutWrapupMiddleware for all three graphs
- #1697: lazy-load __init__.py for agent.middleware, agent.tools,
agent.dashboard (PEP 562); defer heavy imports (exa_py in web_search,
agent.webapp in request_pr_review, deepagents in sandbox.py); add
ttl_cache.py with stale-while-revalidate for tool loaders
Refs: #137
* fix: restore login page render and clear CI lint/format
The plan-review port removed the authRedirectUrl import from login.tsx
but left its call site, crashing the login page at runtime (blank page,
no 'Sign in to open-swe'). Pass the relative path straight to loginUrl,
matching the plan route and the backend relative-redirect handling.
Also drop an unused os import in the guard test and reformat
workflow_push_guard.py to satisfy ruff.
* fix: restore RepairOrphaned middleware export and repoint model fake to deferred_model boundary
* fix: restore RepairOrphanedToolCallsMiddleware, fix E2E model-fake patch, drop dead ttl_cache
- Re-add RepairOrphanedToolCallsMiddleware to the lazy middleware __init__
(_MIDDLEWARE_MODULES, __all__, TYPE_CHECKING) so agent.reviewer can import it.
- Reroute E2E model patching to deferred_model.make_model so make_model_or_defer
(used by all three graph factories) returns the scripted fake instead of
building a real model with fake credentials.
- Drop unused agent/utils/ttl_cache.py — no agent module imports it.
- Fix import ordering in agent/reviewer.py and agent/analyzer.py (ruff I001).
- Format tests/test_dispatch.py.
* fix: claim-then-post run-level failure dedup; stop permanent suppression
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
* feat: port plan-review & workflow-approval UX (#135)
Port six upstream commits onto dev:
- c03a6be7 (already ported): keep plan guidance high-level
- 546042a4: add workflow approval UI with diff preview, approval URLs,
web review links, and polling for approval status during active runs
- 216cf181: remove workflow token elevation; approved pushes pass
through directly without proxy token rewriting
- 3dbc0282: preserve plan redirects after login by accepting relative
same-origin redirect_to values and rejecting blocked paths
- bb104d93: submit plan comments with cmd+enter
- 90cb6caa: terse Slack replies, shared content via save_plan outside
plan mode (PLAN_STATUS_SHARED), reject shared-content mutations
Refs: #135
* fix: restore login page render and clear CI lint/format
The plan-review port removed the authRedirectUrl import from login.tsx
but left its call site, crashing the login page at runtime (blank page,
no 'Sign in to open-swe'). Pass the relative path straight to loginUrl,
matching the plan route and the backend relative-redirect handling.
Also drop an unused os import in the guard test and reformat
workflow_push_guard.py to satisfy ruff.
* fix: carry workflows:write on the standing proxy token
Complete the half-ported upstream 216cf181 cascade. The port dropped
_run_with_workflow_token from the guard but missed the paired github_app
change, so an approved .github/workflows push ran with the base token
(no workflows:write) and GitHub 403'd it.
Add workflows:write to BASE_RUNTIME_PROXY_TOKEN_PERMISSIONS and delete the
now-orphaned WORKFLOW_RUNTIME_PROXY_TOKEN_PERMISSIONS constant; update the
github_app and proxy_auth tests to match. The HITL approval gate in
workflow_push_guard.py is unchanged — this only lets the standing token
push once a human approves.
* fix: restore transient workflow-token elevation (revert standing workflows:write)
The standing GitHub-App proxy token (BASE_RUNTIME_PROXY_TOKEN_PERMISSIONS) is
ALWAYS-ON, so carrying workflows:write on it made the fork's HITL workflow-push
guard the sole control over unapproved workflow pushes. The guard's git-push
parser has gaps (obfuscated-expansion push, `gh api` REST contents PUT,
fully-qualified cross-branch refspecs); with a permanently workflows-scoped
token those gaps become live unapproved-workflow-push exploits (1 critical, 2
high — security review BLOCK on #159).
Restore dev's transient-elevation model:
- Drop workflows:write from BASE_RUNTIME_PROXY_TOKEN_PERMISSIONS; re-add the
WORKFLOW_RUNTIME_PROXY_TOKEN_PERMISSIONS constant (base + workflows:write).
- Re-introduce _run_with_workflow_token in the guard: it mints the
workflows-scoped token via refresh_proxy_token around the approved,
guard-normalized fixed_command, then downscopes to RUNTIME then BASE in a
finally. Route the approval branch through it.
- Restore the dev token/elevation tests.
The standing token no longer carries workflows:write, so the three parser
bypasses hit GitHub 403 again; an approved push still succeeds because the
elevation grants workflows:write only around the normalized command. Keeps all
of #159's diff-preview / approval-URL / Slack-card guard additions.
* fix: reject protocol-relative path from sanitizeAuthRedirect (open redirect)
sanitizeAuthRedirect returned parsed.pathname+search+hash, which `new URL` can
resolve to a protocol-relative `//host` (e.g. input `/..//evil.com` normalizes
same-origin, passing the origin check, but yields a path starting with `//`).
ClientRedirect / login.tsx feed that path to window.location.replace, so it
navigates cross-origin — an open redirect. Reject any resolved path that is not
a single-leading-slash path (`^/[^/]`), falling back to the default. Adds
coverage for `/..//evil.com`, `/.//evil.com`, and `//evil.com`.
* fix: log SECURITY error when workflow-token downscope fails
The elevate->push->downscope finally block was silent on failure. If both
refresh_proxy_token calls fail, the sandbox retains workflows:write for the
rest of the run with no signal. Log a SECURITY error on the partial and full
downscope-failure paths so the retention is observable.
Addresses the GPT-4.1 cross-family review of the token-scope remediation.
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
* feat: port LangSmith LLM Gateway routing from upstream (#1671, #1673, #1674, #1678)
Ports four upstream commits that add opt-in LLM call routing through the
LangSmith Gateway, preserving fork conventions (Bedrock/Fireworks model IDs,
no-agent-attribution, bun toolchain).
- #1671 (e9dc6e01): opt-in gateway routing — new gateway.py, team-settings
toggle, admin UI section, wired into make_model for all graph entrypoints
- #1673 (702ef908): dedicated LANGSMITH_GATEWAY_API_KEY precedence over
platform LANGSMITH_API_KEY
- #1674 (5f7c2f46): fix Fireworks gateway base URL to /fireworks (bare host,
SDK appends /v1/chat/completions) + SanitizeFireworksMessagesMiddleware
- #1678 (73b7d1c0): fix OpenAI Responses reasoning replay —
SanitizeOpenAIResponsesMiddleware, store/include config for encrypted
reasoning content, reasoning_effort coercion for Chat Completions fallback
Refs #134
* fix: downgrade gateway not-routed log to debug, add Bedrock UI note, add sanitizer parity
- Downgrade logger.warning to logger.debug in gateway_overrides for
not-routed providers and missing API key (Bedrock is the default
provider in this fork, so these are expected steady states)
- Add Bedrock to the LLMGatewaySection route-toggle description so
admins know it is not routed through the gateway
- Add SanitizeOpenAIResponsesMiddleware to chat.py for parity with
server.py and reviewer.py
- Restore the Bedrock region comment in model.py that explains the
AWS_REGION / AWS_DEFAULT_REGION precedence
Refs #138
---------
Co-authored-by: amoussa1229 <166072409+amoussa1229@users.noreply.github.com>
A sweep of the deferred backlog found two rows whose changes already shipped
in dev via #81 (the upstream-sync merge) — an empty cherry-pick confirms no net
change:
- #1618 (omit plan link in PR description when no plan exists) — was 'likely regression'
- #1646 (install missing deps before verification)
No code change; ledger accuracy only.
* feat: auto-load scoped AGENTS on reads (#1684)
Adapted from upstream langchain-ai/open-swe #1684 to the fork's
direct-import middleware registry and middleware stack ordering.
Adds SubdirAgentsReadMiddleware, which appends applicable ancestor
AGENTS.md instructions to read_file results once per run, so scoped
rules are visible before edits. Wired into get_agent immediately after
ToolErrorMiddleware, matching upstream's relative position.
Note: this changes file-read behavior for every main-agent run. The
reviewer graph uses its own leaner middleware stack and is unaffected.
(cherry picked from commit 7f7af71547be2199cea699284676f8ceefba7691)
* feat: add platform issue reporting tool (#1685)
Adapted from upstream langchain-ai/open-swe #1685 to the fork's
direct-import tool registry. Adds the report_platform_issue tool
(stdlib-only: returns a locally generated UUIDv7 report id, no external
network call) and wires it into get_agent's curated tool list.
Dropped upstream's test_task_retry_wraps_inside_tool_error_middleware
assertion, which references ToolRetryMiddleware that this fork does not
wire into the middleware stack.
(cherry picked from commit 88b62322b44103335773002d0745704cd96e9160)
---------
Co-authored-by: Ramon Nogueira <ramon.nogueira@langchain.dev>
* feat: editable plan mode — owner hand-edits the plan before approval (#1610, #80)
Adds a PUT /dashboard/api/plan/{thread_id} endpoint and PlanReview UI edit
mode so the thread owner can refine the published plan markdown by hand. The
edited markdown is re-published as "ready" (preserving reviewer comments),
mirrored into the sandbox plan.md, and handed to the agent as the source of
truth on approve.
Approve now reads the published plan content strictly (raise_on_error=True)
so a transient store failure aborts instead of silently dropping an edited
plan, matching the comment-read contract.
The banner-overlap fix (collapsed git-panel clearing the "Review plan" link)
was already ported in #128; this picks up the remaining edit-mode pieces.
Refs #80
* fix(plan): make approve_plan idempotent, dispatch before persisting, fix comment count
SH-128-03: approve_plan set status APPROVED before dispatching the follow-up
run and had no already-approved guard, so a failed dispatch left the plan stuck
approved-but-undispatched and a double-submit double-dispatched + double-posted
the Slack notice. And the Slack notice counted len(comments) including empty
comments _format_comments filters out.
- Return 409 when the plan is already approved (idempotent double-click/retry).
- Dispatch the implementation run BEFORE persisting APPROVED so a dispatch
failure leaves the plan re-approvable. _dispatch_followup passes plan_mode
explicitly, so the run is unaffected by the reorder.
- Count only non-empty comments in the Slack approval notice.
Fixed here (not on #128) because #128's approve_plan is rewritten on this
branch; #129 inherits it. Adds tests for the 409, the filtered count, and the
dispatch-before-status ordering.
---------
Co-authored-by: seahaven-openswe[bot] <296972425+seahaven-openswe[bot]@users.noreply.github.com>
Co-authored-by: Adam Moussa <adam@seahavenind.com>
