deps-dev(deps-dev): bump the patch-and-minor group across 1 directory with 2 updates

[dependabot]

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

Updates @langchain/langgraph from 1.4.9 to 1.4.13

Release notes

Sourced from @​langchain/langgraph's releases.

@​langchain/langgraph@​1.4.13

Patch Changes

• Updated dependencies [f8bdf16, cef10ab]:
  • @​langchain/langgraph-sdk@​1.10.0

@​langchain/langgraph@​1.4.12

Patch Changes

• #2714 a2a59ec Thanks @​hntrl! - Update checkpoint integrations to require the patched checkpoint serializer release.

• Updated dependencies [a2a59ec]:

  • @​langchain/langgraph-checkpoint@​1.1.5

@​langchain/langgraph@​1.4.11

Patch Changes

• #2706 eaa5472 Thanks @​zduric-langchain! - fix(langgraph): dedupe merged callback handlers by identity

  mergeCallbacks concatenated handlers and inheritableHandlers while deduping tags, so a handler inherited by both the ambient and the explicit config picked up an extra registration at every graph boundary. With tracing on, a nested streamMode: "messages" run delivered every token twice.

• Updated dependencies [3ce9f8d, 51b4202, a86f813]:

  • @​langchain/langgraph-sdk@​1.9.30

@​langchain/langgraph@​1.4.10

Patch Changes

• #2690 00f68a1 Thanks @​saad-supports-langchain! - fix(langgraph): keep context values out of tracer-derived metadata

Changelog

Sourced from @​langchain/langgraph's changelog.

1.4.13

Patch Changes

• Updated dependencies [f8bdf16, cef10ab]:
  • @​langchain/langgraph-sdk@​1.10.0

1.4.12

Patch Changes

• #2714 a2a59ec Thanks @​hntrl! - Update checkpoint integrations to require the patched checkpoint serializer release.

• Updated dependencies [a2a59ec]:

  • @​langchain/langgraph-checkpoint@​1.1.5

1.4.11

Patch Changes

• #2706 eaa5472 Thanks @​zduric-langchain! - fix(langgraph): dedupe merged callback handlers by identity

  mergeCallbacks concatenated handlers and inheritableHandlers while deduping tags, so a handler inherited by both the ambient and the explicit config picked up an extra registration at every graph boundary. With tracing on, a nested streamMode: "messages" run delivered every token twice.

• Updated dependencies [3ce9f8d, 51b4202, a86f813]:

  • @​langchain/langgraph-sdk@​1.9.30

1.4.10

Patch Changes

• #2690 00f68a1 Thanks @​saad-supports-langchain! - fix(langgraph): keep context values out of tracer-derived metadata

Commits

• 3609b35 docs: fix typo in cache policy comment (#2740)
• 3ed60cb chore: version packages (#2728)
• 6530ba9 chore: version packages (#2715)
• c3b27a9 fix(checkpoint): narrow re-constructable types in JsonPlusSerializer (#2709)
• 659d628 chore: version packages (#2704)
• eaa5472 fix(langgraph): dedupe merged callback handlers by identity (#2706)
• f6b41bf chore: version packages (#2698)
• 00f68a1 fix(langgraph): keep context values out of tracer-derived metadata (#2690)
• 6ac60da chore(deps): bump the langchain group across 1 directory with 5 updates (#2651)
• See full diff in compare view

Updates darwin-agents from 0.15.0 to 0.16.0

Release notes

Sourced from darwin-agents's releases.

v0.16.0: empirical Bernstein confidence sequence

v0.15 measured its own two sequential methods honestly and left the conclusion on the record: mSPRT does not hold its configured alpha at Darwin's sample sizes, Hoeffding does but is too conservative to resolve realistic gaps, and the proper fix is "an unknown-variance e-process or a t-mixture, which is a different method, not a patch". This release ships that method.

confidenceMethod: 'eb'

The predictable plug-in empirical Bernstein confidence sequence of Waudby-Smith & Ramdas (JRSS-B 2024, Theorem 2). Time-uniform at level alpha by a nonnegative-supermartingale argument plus Ville's inequality, with no i.i.d. assumption and no variance plug-in inside the guarantee: the bet adapts to the estimated spread, and since any predictable bet is valid, that estimate can only change the power, never the level. That is the structural difference from mSPRT, where the estimate sits inside the guarantee itself.

Wired through SafetyGate (full alpha, no fallback needed), evolution.safety, darwin evolve --confidence-method eb, and persisted evolution flags. ebTwoSample and ebIntervalForArm are exported from the package root.

Measured, not claimed

Decision points pinned by tests/sequential-eb.test.ts so they cannot rot:

Scenario	'eb'	'hoeffding'
constant arms 0.10 vs 0.95	n = 21/arm	n = 32
sigma about 0.05, gap 0.30	n about 59/arm	n = 359
sigma about 0.05, gap 0.20	n about 89/arm	n = 900
judge noise sigma about 0.10, gap 0.10	n about 188/arm	n = 4216

Structural blind zone through n = 17 per arm on a [0,1] score, surfaced via inconclusiveByConstruction. The roughly 0.009 composite deltas our own fleet produces remain out of reach for every method here, EB included: the L/(c*n) floor does not shrink with the variance. What EB changes is that gaps in the 0.1 to 0.3 range become resolvable inside a real test's lifetime.

The imported inequality is checked, not believed

Unlike the Hoeffding boundary, whose proof fits in four lines in its own docstring, EB cites its supermartingale inequality from the paper. The test suite compensates three ways: the inequality is evaluated as an exact finite sum over a grid of discrete laws, bets and predictable means; the empirical type-I error under continuous peeking (the production access pattern) is measured and stays at or below alpha, in contrast to mSPRT's measured drift; and the implementation is pinned to 1e-12 against an independent transcription of the paper's formulas. Five mutation probes each turned tests red, so none of the new tests is vacuous.

Adversarial verify

Round 1 came back NO-GO with three findings, all of the exact v0.15 class where text claims something the code does not do: per-arm versus total units in the inconclusive reason string, ebIntervalForArm silently defaulting a non-finite score bound that the entry point refuses, and the CLI help not listing the new method. All fixed, with a regression test pinning guard parity. Round 2 GO, with the statistical core surviving every kill attempt (independent third transcription of the formulas, a denser supermartingale scan, and a type-I Monte Carlo run with an independent PRNG).

Also

• npm run test:coverage:lcov writes an lcov.info via the built-in reporter, for machine consumers.
• 776 tests, 775 pass, 1 pre-existing environment-gated skip.

Full detail in CHANGELOG.md and the Statistical scope section of the README.

Changelog

Sourced from darwin-agents's changelog.

[0.16.0] - 2026-08-15

v0.15 measured its own two sequential methods honestly and left the conclusion on the record: mSPRT does not hold its configured α at Darwin's sample sizes, Hoeffding does but is too conservative to resolve realistic gaps, and the proper fix is "an unknown-variance e-process or a t-mixture, which is a different method, not a patch". This release ships that method.

Added

• confidenceMethod: 'eb': the predictable plug-in empirical Bernstein confidence sequence of Waudby-Smith & Ramdas (JRSS-B 2024, Theorem 2; arXiv:2010.09686), as ebTwoSample / ebIntervalForArm in src/evolution/sequential.ts, both exported from the package root under the same transparency contract as hoeffdingHalfWidth. Time-uniform at level α by a nonnegative-supermartingale argument plus Ville's inequality. No i.i.d. assumption; no variance plug-in inside the guarantee (the bet λ_t adapts to the estimated spread, and since ANY predictable bet is valid, the estimate can only change power, never level: the structural difference from mSPRT). Wired through SafetyGate (full α, no fallback needed: the regularised variance estimate exists from the first observation, so there is no no-spread abstention), evolution.safety, darwin evolve --confidence-method eb, and persisted evolution flags.
• Measured decision points, pinned by tests/sequential-eb.test.ts so they cannot rot (EB exact for constant arms, median over 21 seeded runs for noisy ones; Hoeffding columns are the exact first n at which its data-independent bar drops below the gap): constant 0.10 vs 0.95 at n=21 (Hoeffding 32); σ≈0.05 gap 0.30 at n≈59 (Hoeffding 359); gap 0.20 at n≈89 (Hoeffding 900); judge-noise σ≈0.10 gap 0.10 at n≈188 (Hoeffding 4216). Structural blind zone through n=17 per arm at defaults, surfaced via inconclusiveByConstruction exactly like Hoeffding's. The ~0.009 composite deltas our own fleet produces remain out of reach for every method; the README table says so.
• The imported inequality is checked, not believed: the supermartingale bound is cited from the paper rather than re-derived, so the test suite (a) evaluates E[exp{λ(Y−μ) − 4(Y−m̂)²ψ_E(λ)}] as an EXACT finite sum over a grid of discrete laws, bets and predictable means and asserts it never exceeds 1, (b) measures the empirical type-I error under continuous peeking (the production access pattern) and asserts it stays at or below α at every horizon tried, in contrast to mSPRT's measured drift, and (c) pins the implementation against an independent reference transcription of the paper's formulas to 1e-12. Five mutation probes run during development (α-split dropped, ψ_E scale dropped, samples sorted, structural floor removed, gate dispatch removed) each turned tests red; none of the new tests is vacuous.
• Order sensitivity documented as a contract: the bet is predictable from the prefix, so the same multiset in a different order legitimately produces a different (equally valid) interval. The docstring says why, the gate feeds chronological samples (tracker.getCompositeScores), and a pinned test points anyone who "fixes" it by sorting at the docstring.

... (truncated)

Commits

• 02c0380 feat(v0.16.0): empirical Bernstein confidence sequence — the unknown-variance...
• See full diff in compare view

[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.