deps-dev(deps-dev): bump the patch-and-minor group with 2 updates

[dependabot]

Bumps the patch-and-minor group with 2 updates: @langchain/langgraph and darwin-agents.

Updates @langchain/langgraph from 1.4.7 to 1.4.8

Release notes

Sourced from @​langchain/langgraph's releases.

@​langchain/langgraph@​1.4.8

Patch Changes

• #2604 42ec394 Thanks @​hntrl! - fix(pregel): wait for completed-superstep persistence with sync durability

• Updated dependencies [f326b89, c201256]:

  • @​langchain/langgraph-sdk@​1.9.26

Changelog

Sourced from @​langchain/langgraph's changelog.

1.4.8

Patch Changes

• #2604 42ec394 Thanks @​hntrl! - fix(pregel): wait for completed-superstep persistence with sync durability

• Updated dependencies [f326b89, c201256]:

  • @​langchain/langgraph-sdk@​1.9.26

Commits

• 3dccad1 chore: version packages (#2603)
• 42ec394 fix(core): enforce sync durability barriers (#2604)
• 644892c docs(langgraph): fix README logo paths (#2599)
• 8fb6376 chore: update Vitest to 4.1.9 (#2587)
• 81f6365 chore(deps): bump the langchain group with 8 updates (#2578)
• See full diff in compare view

Updates darwin-agents from 0.11.0 to 0.12.2

Release notes

Sourced from darwin-agents's releases.

v0.12.2 — evolution optimizer/reflector drop bypassPermissions (security)

Security patch. buildEvolutionLoop's optimizer + GEPA-reflector subprocesses are pure text mutators but ran the Claude CLI with --permission-mode bypassPermissions and no --allowedTools restriction. Their input quotes untrusted agent output (critic feedback can contain scraped web content), so a prompt injection could in principle have steered an unrestricted subprocess into tool calls. Both closures now run in the CLI's deny-by-default permission mode (autonomous: false) — legitimate runs are unaffected. Regression tripwire: tests/build-loop-security.test.ts. 625 tests.

v0.12.1 — loop-composition surface exported from the root

Export patch, no behavioural change. buildEvolutionLoop + DarwinLoop/ExperimentTracker/PatternDetector/PromptOptimizer/SafetyGate (+ EvolutionResult/AgentToolContext types) are now importable from the package root — completing the v0.12.0 bring-your-own-judges story: external post-run hooks can now score with custom judges (criticPrompts) and drive the same evolution cycle, without deep dist/ imports (which the exports map blocks). buildEvolutionLoop mirrors the CLI run-path wiring incl. opt-in GEPA via agent.evolution.useGepa. 623 tests.

v0.12.0 — bring your own judges + adapter peer-compat release guard

Bring-your-own-judges: the multi-critic runner accepts caller-supplied critic sets, closing the gap that forced fleets with domain agents to fork this file. Plus a release guard against the adapter peer-range breakage that has now happened twice. Default behaviour is unchanged.

Added

• `RunMultiCriticOptions.criticPrompts` — an explicit critic-prompt set per `runMultiCritic` call, bypassing the built-in `getCriticPrompts` name lookup. The built-in map covers a handful of generic archetypes; any other agent name silently fell back to the investigator judges, mis-scoring domain output — the only workaround was forking `multi-critic.ts`. Any count ≥ 1 works; invalid entries are dropped; an empty/non-array/all-invalid value falls back to the built-in lookup (never zero judges). Judge contract: each prompt must instruct the critic to emit `===SCORE=== N` (or an `X/10` figure).
• `RunMultiCriticOptions.outputLabel` — overrides the evaluation-preamble label for agents outside the built-in `AGENT_OUTPUT_LABELS` map. Whitespace-only values are ignored. Composes with `normalizeForJudging` and `criticPrompts`.
• Release guard `check:adapter-compat` (wired into both CI and `prepublishOnly`): fails the build when this package's version escapes the published `darwin-langgraph@latest` peer range. Both 0.9.0 and 0.11.0 escaped the adapter's then-current cap the day they shipped — explicit paired installs failed with `ERESOLVE`, unpinned installs silently downgraded darwin-agents. Prerelease-aware (judges the release counterpart), network-tolerant (registry hiccups warn and pass). Pairs with `darwin-langgraph@0.5.4`, whose peer now accepts any darwin-agents 0.x.

Notes

• 620 tests (619 pass / 1 pre-existing skip). No behavioural change without the new options.
• `semver` + `@types/semver` are devDependencies for the guard script only — the runtime stays zero-dependency.

Changelog

Sourced from darwin-agents's changelog.

[0.12.2] — 2026-07-17

Security

• Evolution-loop optimizer/reflector subprocesses no longer run with bypassPermissions. buildEvolutionLoop's two LLM closures (legacy optimizer meta-prompt + GEPA reflector) are pure text mutators, but they passed autonomous: true to runAgent, which spawns the Claude CLI with --permission-mode bypassPermissions — and since neither agent definition declares tools/MCP servers, no --allowedTools restriction was emitted either. Their input quotes untrusted agent output (critic feedback reports can contain scraped web content), so a prompt injection could in principle have steered an unrestricted subprocess into tool calls. Both closures now run autonomous: false (the CLI's deny-by-default permission mode); legitimate runs are unaffected because the templates demand "return ONLY the prompt text". Found by an adversarial review of the first external consumer wiring buildEvolutionLoop (severity: plausible, low probability, real surface). tests/build-loop-security.test.ts is the regression tripwire.

[0.12.1] — 2026-07-16

Export patch — no behavioural change. Completes the v0.12.0 "bring-your-own-judges" story: custom judges were usable, but an external post-run hook still could not drive the evolution loop from the published package, because the loop-composition surface was internal and the package exports map (correctly) blocks deep imports from dist/.

Added

• Root exports for the loop-composition surface: buildEvolutionLoop (wires tracker/patterns/safety/legacy-optimizer/opt-in-GEPA/notifications around a DarwinLoop exactly like the CLI run path — GEPA activates via agent.evolution.useGepa), plus the individual classes DarwinLoop, ExperimentTracker, PatternDetector, PromptOptimizer, SafetyGate and the types EvolutionResult / AgentToolContext for consumers composing custom loops. New tests/root-exports.test.ts locks the surface.

[0.12.0] — 2026-07-16

Bring-your-own-judges: the multi-critic runner accepts caller-supplied critic sets, closing the gap that forced fleets with domain agents to fork this file. Plus a CI release-guard against the adapter peer-range breakage that has now happened twice. Default behaviour is unchanged.

Added

• RunMultiCriticOptions.criticPrompts — an explicit critic-prompt set for a single runMultiCritic call, bypassing the built-in getCriticPrompts name lookup. The built-in AGENT_CRITIC_MAP covers a

... (truncated)

Commits

• 5c8a908 test(v0.12.2): match only code occurrences of autonomous: false
• 19496ef fix(v0.12.2): evolution optimizer/reflector drop bypassPermissions
• 24b8f03 feat(v0.12.1): export the loop-composition surface from the package root
• 65be1c5 feat(v0.12.0): bring-your-own-judges + adapter peer-compat release guard
• 2f9d677 deps-dev(deps-dev): bump the patch-and-minor group with 2 updates (#18)
• See full diff in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore <dependency name> major version will close this group update PR and stop Dependabot creating any more for the specific dependency's major version (unless you unignore this specific dependency's major version or upgrade to it yourself)
• @dependabot ignore <dependency name> minor version will close this group update PR and stop Dependabot creating any more for the specific dependency's minor version (unless you unignore this specific dependency's minor version or upgrade to it yourself)
• @dependabot ignore <dependency name> will close this group update PR and stop Dependabot creating any more for the specific dependency (unless you unignore this specific dependency or upgrade to it yourself)
• @dependabot unignore <dependency name> will remove all of the ignore conditions of the specified dependency
• @dependabot unignore <dependency name> <ignore condition> will remove the ignore condition of the specified dependency and ignore conditions

[dependabot]

Labels

The following labels could not be found: dependencies. Please create it before Dependabot can add it to a pull request.

Please fix the above issues or remove invalid values from dependabot.yml.