Skip to content

v4.0.0

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 16:07
· 26 commits to main since this release
48c73d2

Bernstein 4.0.0

At a glance

  • If a workflow node fails, Bernstein skips later layers, including independent branches, as in bernstein workflow resume. A detached SSH task with a nonzero remote exit fails the run and leaves later tasks pending.
  • bernstein govern audit --strict gives CI distinct exit codes for failed checks and unmeasurable prerequisites.
  • At run close, Bernstein seals the final delegation digest, hop count and manifest configuration with the install audit key.
  • Fleet selection identifies an enabled instance where one worker matches the task type, dispatch tier and required capabilities.

What is partial in 4.0.0 summarizes selected features and their limits.

Before you upgrade

Read the 3.24 to 4.0 migration guide.

  1. Quality gates. Previously ignored settings such as quality_gates.large_file_check in bernstein.yaml now apply. A value of the wrong type raises SeedError (#6615).
  2. Delegation verification. If a manifest declares a final head with a missing or invalid seal, bernstein delegation verify exits 1. This includes heads declared before seals existed. An otherwise intact, nonempty chain with no final declaration reports valid=false and exits 3. An interrupted close can leave the declaration missing (#5352).
  3. MCP versions. Bernstein uses SDK 2.3.x with the 2025-11-25 task wire contract. By default, the HTTP transport returns 400 for an unsupported MCP-Protocol-Version header. During the handshake, the server answers an unknown initialize.params.protocolVersion with its default revision.
  4. Quarantine. The default workspace skill loader saves initialization failure reasons across restarts. A bare SkillLoader keeps them in memory. Persisting runtime plugin failures in quarantine requires a workspace store and record_declaration_failures=True.

After upgrading, run the audit from the project root and check which prerequisites it could measure.

bernstein govern audit --format json --strict

Governance and audit

bernstein govern audit writes text, JSON or SARIF 2.1.0. Run it from the repository root for workspace-relative SARIF locations. --quiet suppresses the report.

Result Default --strict
Any check fails or crashes, or a measured finding has passed=null 1 1
Otherwise, at least one prerequisite is unmeasurable 1 2
A nonempty selection returns only passes, declarations or both 0 0
Empty selection 1 1
CLI usage error 2 2

The full table is in governance operations.

  • Compare reports. Use --since BASE_SHA256 --against HEAD_SHA256 together to compare two stored audit reports offline.
  • Policy-library checks. govern audit --only CMP runs the checks reported in compliance check JSON.
  • ISO/IEC 42005 export. bernstein audit export --standard iso-42005 exports the evidence Bernstein holds for an AI impact assessment.
  • Control coverage. bernstein compliance controls lists the benchmark suites that declare each control and their recorded measurements. Later submissions take precedence. The display says unavailable when no measurement exists.
  • Scorecard baselines. An install-identity baseline stays neutral without a matching trusted key. Supply the public key with bernstein bench run SUITE --baseline BASELINE.json --trusted-key FINGERPRINT=PATH. Stub baselines require --stub-signer and successful replay.

Evidence integrity and supply chain

  • Delegation seals. Verification detects records removed, replaced or appended at the tail after close. It cannot detect replay of an authentic snapshot, an unrecorded action or changes made with the audit key. Older runs receive no backfill.
  • Hash profile. hash_profile: jcs-v2 in bernstein.yaml opts a new run's journal, lineage spine and receipt into RFC 8785 canonical JSON. Bernstein signs the receipt only with a signing key configured. py-json-v1 remains the default.
  • Release provenance. GitHub Releases include a Sigstore provenance bundle.
  • OSV-Scanner. The adapter scans against a pinned local database and rejects digest mismatches and time-dependent rules (ignoreUntil, effectiveUntil). This covers part of #2958.
  • CoSAI evidence map. The map includes MCP-T1 through MCP-T12 from the 2026-01-08 revision.
  • Custody signers. Detached-JWS lineage signers can use an existing custody signer without exporting the private key.
  • Offline signing. Offline wheelhouse signing keeps working with cosign 3.1.4.

Fleet routing and resources

  • Dispatch tiers. Generated tasks get a dispatch tier from their runtime requirements. Each batch runs tasks from one tier. Tiers survive REST requests, PostgreSQL storage, YAML plan loading and both retry paths.
  • Routing hints. The model, effort, cli and requires hints survive a restart in the REST API and the JSONL TaskStore.
  • Worker heartbeats. Disk, memory and mesh-latency readings can guide node selection, task stealing and worker pulls.
  • Live view. bernstein live shows registered remote workers and the tasks they claimed.

Execution correctness

  • Cost corrections. The budget stop counts increases in explicit cumulative cost above spend already recorded, even when token counts stay flat. The current snapshot format omits the prior explicit cost. After restart, the first report for a restored entry can miss a cost increase if no tokens were added.
  • Orphaned tasks. After a clean worker exit, an eligible non-artifact task with no new commits or worktree changes can complete if its declared completion signals already pass on the base commit.
  • Completion timeout. BERNSTEIN_SIGNAL_TIMEOUT_S sets the timeout for a test_passes command. It defaults to 120 seconds.
  • Sandbox leases. A holder can release only its own active lease. bernstein doctor reports owner task statuses for confirmed leaks older than the configured threshold (24 hours by default). It warns when a confirmed leak's age is unknown.
  • Reconcile helper. It skips completed work, stops on failure and can resume a trusted partial run.
  • Required gates. A failing or inconclusive plugin result blocks promotion when GatePipelineStep.required is true. An explicit block from the plugin also applies on an optional step.

Models and adapters

  • OpenAI and Codex. provider=openai uses the OpenAI API client. The Codex adapter validates and forwards reasoning effort before spawning.
  • Retries. An explicit role or run model pin takes precedence. Otherwise, a retry keeps a supplied provider-prefixed model. With neither a known model nor a pin, the adapter resolves its default (#6605).
  • Pi. Pi runs with --print --mode json and streams events to the runner log.
  • New adapter contracts. The release adds CLI contracts for cline, autohand, openhands, cody, composio and ralphex. Adapters generated from the template match the current CLIAdapter contract.

MCP and spawn policy

  • MCP tasks. The persistent TaskStore backs task creation, polling, result retrieval, listing and cancellation on the 2025-11-25 wire contract.
  • Capability observations. With --observe-capabilities, the gateway hashes each complete tool definition and each top-level field after the upstream handshake. The snapshot stores those digests. On a change, mcp.capability_drift records previous and current digests and the changed field names per tool. The repository includes test vectors.
  • Server manifest verification. With signing checks enabled and an audit chain supplied, the MCP pre-load check records the server manifest's signature verdict, the bundle scan status and the load decision.
  • External policy registry. For fresh direct-adapter spawns, an injected registry checks each attempt's adapter, model and endpoint, including failover. At least one policy hook must return ALLOW.
  • Role allow-lists. Roles without a named policy can use a * adapter allow-list. An empty default denies every adapter.

What is partial in 4.0.0

Component Works in 4.0.0 Not included
Graph trace ingest (#4964) Maps recorded structure to RunPlan / PlanNode and hashes tool-call arguments Scheduling, runtime callback conformance, CLI, replay and offline receipts
Plan graph fold (#3614) Rebuilds the recorded plan graph offline from the journal Scheduler integration, amendments
Agent Manifest export Composition data, signed availability companion and per-hop artifact references A complete verdict
Registry federation Opt-in for Python callers, Ed25519 pins, unsigned indexes refused by default Donor CLI configuration
Attachment manifests A read-only preview returns copies of selected existing tasks and the manifest digest Creating or persisting derived tasks
External policy registry (#4912) Checks fresh direct-adapter attempts, including failover CLI policy loading and other execution paths
Worker telemetry Node selection, task stealing, worker pulls Orchestrator claims by task ID (#7853)
Registry partitions Partitions inside one process HTTP sharding
Operator halt controller Pauses and resumes new TaskStore claims Stopping running workers, installation-wide halt
Retry model retention (#6605) Keeps a supplied provider-prefixed model unless a role or run pin overrides it Resolving the adapter default in the retry request

Dashboard, diagnostics and developer experience

  • Dashboard. Task status shows tenant-scoped coordinator readiness. MCP status reports the coordinator and bridge versions separately.
  • Quarantine. Workspace quarantine can also record reasons for malformed routine trigger declarations and invalid issue trackers.
  • Windows. Artifact-mode completion handles agent IDs containing :. A directory at a declared output path reports "wrote no output".
  • Telemetry. bernstein telemetry verify-span reads span JSON from standard input with --span - or --span @-.
  • Snapshots work inside linked Git worktrees. Flaky-test detection retains pytest errors, and GitLab webhook events follow configured trigger rules.

Documentation, CI and maintenance

  • Workflow docs. The manifest reference distinguishes dispatch, verified completion and successful command execution. Resume examples supply the goal again because snapshots omit it.
  • CI. Dependabot skips auto-merge on draft PRs. Maintainer review on protected automation changes excludes known writers. A maintainer PR follows ordinary review rules if a commit author is neither on the roster nor the built-in github-actions[bot] account.
  • Characterization tests. Strict expected failures document known gaps in workflow execution and recovery (#6639).

Known limitations

  • With an injected policy registry, bridge, in-process, container, sandbox and crash-resume paths refuse before execution. CLI policy loading is still pending (#4912).
  • The detached local path without an SSH sidecar records task lifecycle events but does not yet execute the submitted task.
  • Regular fixes ship on 4.0.x. The only planned 3.24.x backport is 3.24.2.

The complete list of changes follows below and is in docs/release-notes/v4.0.0.md.

Full list of changes

Added

  • Validate attachment manifests and preview their derived tasks without writing to the task store.
    (#6912)
  • Define an optional sandbox backend contract for verified reattachment and removal.
    (#6911)
  • Include tenant-scoped coordinator readiness observations in dashboard task status.
    (#7022)
  • Compare two stored governance audit reports offline with govern audit --since and --against.
    (#7029)
  • Let MCP gateways explicitly opt in to recording capability observations.
    (#7036)
  • Export ISO/IEC 42005 impact-assessment evidence with bernstein audit export --standard iso-42005.
    (#7085)
  • Find the first changed or missing property in a pair of receipt render captures.
    (#7101)
  • Aggregate supplied run-conduct projections within a specified time window.
    (#7102)
  • Attach the Sigstore provenance bundle to GitHub Releases.
    (#6968)
  • Provide an opt-in failure sentinel for testing audit detection, recording, and notification.
    (#6916)

Changed

  • Select fleet instances based on explicit worker capabilities rather than generic matches.
    (#6910)
  • Update shipped Web UI dashboard assets with React Query 5.105.1.
    (#6933)
  • Expose doctor installation checks as governance-audit report producers.
    (#6918)
  • Migrate the quality-gate template to the standardized GatePlugin interface.
    (#6964)
  • Quarantine skills that fail initialization with a journaled reason instead of failing silently or
    crashing. (#6917)
  • Capture MCP server capability snapshots only after completing the upstream handshake.
    (#7016)
  • Report the coordinator version separately from the bridge version in MCP status.
    (#7023)
  • Make the declaration quarantine check callable through the audit check registry.
    (#7033)
  • Classify and assign dispatch tiers to generated tasks based on runtime requirements.
    (#7034)
  • Derive reconciliation intent keys and check whether a fresh observation already satisfies the intent.
    (#7040)
  • Cache snapshot metadata directory resolution once per listing for improved performance.
    (#7048)
  • Report both implementation modules when a duplicate audit check is registered.
    (#7051)
  • Update agents dependency to ^0.33.0 in the Cloudflare MCP server template.
    (#6950)

Fixed

  • Show registered remote worker nodes and their claimed tasks in the bernstein live AGENTS panel.
    (#6753)
  • Route provider=openai requests directly through the OpenAI API client.
    (#6884)
  • Validate and forward reasoning effort parameters before spawning tasks in the Codex adapter.
    (#6899)
  • Keep tasks from different dispatch tiers in separate execution batches.
    (#6905)
  • Use the configured task server in role prompts, prompt partials, and the manager task-API skill.
    (#6924)
  • Preserve task dispatch tiers across REST API requests and PostgreSQL persistence.
    (#6908)
  • Make the completion-signal timeout configurable in place of the fixed 120-second limit.
    (#6923)
  • Build a verified removal proposal for one leaked sandbox lifetime.
    (#6909)
  • Record MCP tool-signing gate outcomes in the audit chain.
    (#6913)
  • Include full injected tool definitions per server when calculating spawn prompt token budgets.
    (#6907)
  • Fix datasource query receipt recording on Windows by sanitizing directory paths containing colons in
    agent IDs. (#6928)
  • Omit unsupported sampling parameters when the OpenAI Agents adapter uses GPT-5 mini with the Responses
    API. (#6975)
  • Complete clean orphaned tasks when their verification gates already pass on the base commit.
    (#6900)
  • Ensure standard input streams are read reliably across platforms using Click open_file in telemetry
    commands. (#6989)
  • Prevent releasing other active leases by ensuring sandbox holders release only their own assigned
    leases. (#7001)
  • Let the pipeline enforce the required flag for failures reported by a quality-gate plugin.
    (#7003)
  • Quarantine malformed routine trigger declarations to isolate configuration errors without halting
    execution. (#7007)
  • Journal quarantine events for invalid issue trackers discovered during workspace scanning.
    (#7008)
  • Automatically quarantine runtime plugins that encounter initialization failures.
    (#7009)
  • Apply configured trigger rules when processing incoming GitLab webhook events mapped to tasks.
    (#7010)
  • Preserve MCP capability snapshots when concurrent writers update the same store.
    (#7026)
  • Preserve explicitly declared dispatch tiers defined in YAML plan steps.
    (#7027)
  • Verify that the inventory anchor names the inventory recorded by a governance audit report.
    (#7038)
  • Fix snapshot resolution and creation when operating inside linked Git worktrees.
    (#7045)
  • Reject duplicate tunnel driver registrations to prevent driver conflicts.
    (#7046)
  • Handle shallow registration call stacks gracefully in registry guards without raising inspection
    errors. (#7047)
  • Recover the run outcome before recording the run-quiescence observation.
    (#7049)
  • Preserve pytest error outcomes when parsing test evidence in flaky test detection.
    (#7025)
  • Update vulnerable glob and minimatch dependencies and hash-pin the affected pip installs.
    (#6965)
  • Correct documented CLI commands, flags, and the benchmark Lambda link.
    (#7079)

OSV-Scanner adapter, pinned to a local database

get_scanner("osv-scanner") returns a feed-pinned SCA adapter. It runs
osv-scanner scan source --offline against a local OSV database
(OSV_SCANNER_LOCAL_DB_CACHE_DIRECTORY), never against osv.dev, whose answer
changes as advisories are published. The caller pins the database in
scope.config["db_pin"]. The adapter digests every osv-scanner/<ecosystem>/all.zip
under the root and refuses to scan when the digest differs, as the Trivy
adapter does for trivy.db.

Finding identity includes the vulnerability id, normalized lockfile path,
severity, summary and affected package metadata. Windows drive and UNC file
URIs use the shared path normalizer so checkout locations do not change
in-scope finding hashes. Any osv-scanner.toml ignore list in the target tree
feeds the recorded invocation hash. Include/exclude globs and max-depth scope
limits are refused rather than ignored. Partial implementation of #2958.

SARIF reports that explicitly mark any invocation as unsuccessful are
rejected, including reports with no findings. Absent invocation metadata
and successful invocations preserve the existing finding normalization.

Config identity uses the versioned osv-scanner-config-v2 digest of sorted
target-relative POSIX paths and exact file bytes: moving an ignore list between
package directories changes invocation identity, while moving the entire checkout
does not. Before any vendor process, the adapter parses TOML and rejects explicit
IgnoredVulns.ignoreUntil and PackageOverrides.effectiveUntil fields (including
case variants), because OSV-Scanner evaluates them against the current time.
Timeless configs and temporal words in comments or reasons remain supported;
unreadable or invalid TOML configs are refused rather than left unvalidated.

Doctor environment snapshots remain statically typed

Doctor's adapter lookup narrows an optional environment mapping where it reads
PATH, clearing an advisory mypy error without an ignore or a cast. Explicit
paths, including an empty PATH, and fallback to the ambient environment retain
their existing behavior.

Fixed

  • Qualify the catch-up skip list annotation so the schedule store's list
    method does not shadow the built-in type during mypy checks.

Derive the recorded structural plan graph offline

A pure journal fold, exported as bernstein.core.replay.fold_plan_graph, now
reconstructs structural task nodes using the producer's
canonical graph encoding. Missing latest snapshots return an explicit
unavailable result, and malformed graph evidence fails closed. Generated
producer-prefix tests cover independent append failures and determinism across
processes. Scheduler integration and amendments remain deferred (#3614).

Opt-in registry federation for Python callers

Volunteer registry callers can explicitly enable bounded breadth-first index
resolution with per-URL Ed25519 pins, default unsigned refusal, denial rules,
cycle diagnostics and conflict reporting. Project manifests remain the policy
source. Existing flat-index browse behavior is preserved; donor CLI
configuration for federation remains a follow-up.

Fixed

  • Programmatic spawn admission now consults an injected external policy registry before starting an action, records every hook evaluation in the audit chain, and refuses unavailable engines, missing explicit approval and unsupported obligations. Native admission refusals remain authoritative. CLI policy loading remains a separate integration task. (#4912)
  • Direct-adapter attempts, including failover, are admitted against their actual adapter, model and effective endpoint. Policy or audit refusals cannot trigger another executor. Injected registries refuse unbound bridge, in-process, container/sandbox and crash-resume paths; native-only behavior is unchanged.

Graph trace ingest adapter for graph-shaped agent runtimes

Ingests foreign execution traces from graph-shaped agent runtimes through the ingest adapter plugin boundary without modifying core scheduling logic (#4964).

The adapter retains foreign nodes, edges, input/output state and tool calls in its ingested representation, and maps node identities and edges to Bernstein's lineage task graph vocabulary (RunPlan / PlanNode). Every tool call carries a deterministic digest of its normalized arguments. Graph text renderings remain byte-identical across repeated ingests.

Input/output state must be mappings, edge conditions must be booleans, and node
identifiers and edge endpoints must be non-empty strings. Tool calls require a
non-empty name. Display fields reject control characters and line separators
before rendering. Missing roles remain unassigned. The lineage projection keeps
cycles and flattens conditional edges into dependencies; it describes recorded
structure and is not a schedulable execution plan.

This is the entry-point-registered ingest-boundary slice of #4964. String inputs
are JSON; local files require an explicit Path. Malformed tool-call entries are
rejected instead of silently discarded. The fixture is synthetic, and the
dependency-free callback recorder is tested with hand-driven events. Real-runtime
callback conformance and recorded-fixture provenance, production declaration
collection, CLI/replay wiring, and offline receipts remain follow-up work.

  • Allow reusable lineage detached-JWS signers to use an existing custody signer without exporting private key material. Existing PEM callers and signature bytes remain compatible.

Expose all existing CMP policy-library observations through govern audit --only CMP and include identical workspace findings in compliance check
JSON. Policy snapshots now use project observations unless explicitly
overridden by declaration flags. Framework profiles mark required check IDs
without changing verdicts or asserting conformance.

CMP execution errors retain their check area and use the shared registry's
exception isolation on both JSON surfaces. Under govern audit --strict, a
crashed CMP check returns exit 1 rather than the incompleteness exit 2.

The existing framework profile annotation is also retained in each SARIF
result's properties, without changing the check verdict or command exit code.

Governance audit JSON, SARIF and CI exit codes

bernstein govern audit now supports --format text|json|sarif, --strict
and --quiet. JSON preserves the complete finding fields, including id and
message (also included in doctor --json). SARIF 2.1.0 includes check rules
and workspace-relative file locations; passing checks use kind: pass and
level: none. Unmeasurable results use kind: review and level: none, with
their reason in the message. Evidence locators are omitted, and absolute local
paths in diagnostics are redacted. Run from the repository root for code scanning.
--quiet suppresses reports and check-crash warnings while retaining sentinel
journaling and notifications.

The default gate keeps exit 1 for failures, check crashes and unmeasurable
checks. --strict distinguishes unmeasurable prerequisites with exit 2,
unless a check failed or crashed (exit 1 takes precedence). Measured findings
without an explicit pass fail closed. Click usage errors, including incompatible
output flags, also exit 2. Declared findings no longer fail the gate: they
previously caused exit 1 on main and now allow exit 0 when no other finding
blocks it. An empty selection exits 1. See the complete exit-code table in
governance operations.

Add a callback-based reconcile apply helper that records already-satisfied intents as skips, stops on failure, and resumes trusted same-context partial receipts without repeating successful changes. Retain stable intent keys and require separate policy authorization.

  • Add one hash_profile selection in bernstein.yaml for a new run's journal, lineage spine and receipt (signed when a signing key is configured). jcs-v2 opts into RFC 8785; unknown profiles fail closed.

  • The first release containing this fragment is the declared compatibility release. py-json-v1 remains the default for that whole release. Changing the new-run default to jcs-v2 requires a separate later PR after measured soak; this release does not flip it.

  • Existing evidence retains its recorded profile, or py-json-v1 when absent, and reopens and verifies without rewriting historical bytes.

  • Preserve declared verification scopes in plugin results and cached gate reports, and retain the reason for inconclusive plugin results. Legacy cache entries remain without coverage declarations.

Delegation verification detects changes to a sealed final receipt chain

Run close atomically seals the final delegation digest, hop count and manifest
configuration with the install audit key. Every delegation verifier checks that
seal and the declared head/count: tail deletion, substitution and post-close
growth fail, even if a keyless writer recomputes the manifest checksum. Missing
declarations are incomplete (valid=false, exit 3), including pre-upgrade runs
and interrupted finalization. Existing unsigned head declarations fail seal
verification; they are not silently upgraded from the remaining ledger (#5352).

Open and legacy runs retain their intact, authority-checked delegated ceilings
in the agent registry, marked by incomplete_runs; saved version-1 registry
snapshots remain verifiable. Access reviews exclude incomplete or mismatched
chains from verified sources. Incomplete chains never report a passing verdict.

This does not detect replay of an earlier authentic snapshot, prove that every
runtime action was recorded, or protect against a writer holding the audit key.
Deleting a declaration yields incomplete, never a completeness claim. There is
no automatic backfill: the remaining receipts cannot establish a historical tail.

Compliance control-to-suite coverage and latest measurements

bernstein compliance controls now surfaces the declaring benchmark suites and
latest recorded measurement for each registry control across text, Markdown,
and JSON outputs (#5455, #7098):

  • Built-in benchmark suite control declarations are exposed from a single shared
    mapping in bernstein.eval.bench.bench_cli, paired with tests verifying every
    registered built-in suite declares at least one valid registry control ID.
  • bernstein.compliance.control_coverage joins the central control registry
    with built-in (and optional custom .suite.json) declarations and
    SubmissionBundle records, selecting the latest measurement per suite
    deterministically regardless of input order and reporting unavailable when no
    measurement exists.
  • bernstein compliance controls accepts optional repeatable --bundle and
    --bundle-dir flags (falling back to .sdd/bench / .sdd/benchmarks when
    omitted) while preserving existing control keys and stable control ordering.

Authenticate install-identity benchmark baselines in scorecards

Python callers can pass trusted public keys to evaluate_ci_scorecard so a valid
install-identity baseline can produce a verified success or regression result.
Missing trust, invalid signatures and failed receipt replay remain neutral.
bench run --baseline accepts --trusted-key FINGERPRINT=PATH to supply
those public keys explicitly. Without a matching key the comparison stays
neutral.

Keep the pytest controller outside the memory cap

The test suite applies its address-space guard to test workers and plain
runs, leaving the pytest-xdist controller uncapped. Resource failures and
the reported worker-crash controller abort include guard context. Diagnostic
regressions run without the POSIX-only probe on Windows and preserve custom
caps and inherited Python module paths in child sessions.

Add optional composition-only Agent Manifest export with operator declarations,
a signed hash-bound availability companion, and per-hop artifact references.
Pin agent-manifest 0.12.0; preserve its INCOMPLETE composition-only verdict and
strict runtime artifact verification semantics.

CoSAI packs include the pinned MCP threat taxonomy (#6217)

The existing cosai evidence map adds MCP-T1 through MCP-T12 from the selected
2026-01-08 paper revision. Every new row preserves its exact source provenance
and names bounded gate evidence and remaining gaps: four Partial and eight
Not covered, with no claim that the pre-load gate covers an entire threat class.
The original thirteen principle rows, pack schema and ZIP layout are preserved.

Worker health telemetry

Worker heartbeats accept optional disk, memory and mesh latency readings.
The registry uses them for direct node selection, task stealing and
bernstein worker pulls. Orchestrator claims by task ID remain open in
#7853.

Cluster registry ring

Connect the deterministic hash ring to opt-in local partitions inside one
registry process through cluster.shards and cluster.shard_vnodes. Node
registration, heartbeat and administrative mutations use the owning partition;
defaults and the flat node persistence format remain unchanged. HTTP sharding is not enabled.

Fixed

  • Preserve backend dispatch tiers through both task retry paths. Reopen can explicitly select a validated tier, recording the previous tier, selected tier and reason with the persisted transition; omission preserves the current tier. Automatic failure-based retiering remains outside this change. Refs #6560.

Preserve task routing hints across restart

Task REST responses and the JSONL TaskStore now retain exact model, effort,
cli and requires declarations through single/batch creation, later records
and restart. Invalid capability declarations are rejected before REST
persistence; historical tasks keep their existing defaults. This preserves
routing data without adding fleet selection, dispatch or result acceptance.
Refs #6562.

Doctor reports aged leaked lanes with their evidence (#6572)

The existing read-only leak check now includes each confirmed lane's recorded
owner statuses and resource-age source in text and structured JSON. Both sweep
kinds use one canonical task replay snapshot. The existing 24-hour environment
default and strict older-than comparison remain; unusable ages warn explicitly.
Unreadable evidence or an invalid threshold fails the doctor check instead of
being displayed as a successful check. This adds no removal authority or
provider probe.

Retain supplied provider models across task retries

Both retry task creators preserve an already supplied provider-prefixed model
instead of manufacturing a bare Claude tier. Explicit role/run model pins keep
their priority, and existing effort, retry limits and backoff remain unchanged.
This is a partial fix for #6605: missing effective model context still needs
propagation when neither the task nor an explicit pin identifies the model.

Keep the model override unset in both retry paths when neither a prior task
model nor an operator pin is available. The adapter continues resolving its
configured default, rather than receiving an invented Claude tier. Original
effort overrides are preserved alongside an unresolved model so an effort-only
override cannot bypass the original template resolution. Known-model effort
escalation, explicit pins and resolved provider-model retention remain unchanged.

The seed parser now reads every quality_gates.* key that QualityGatesConfig defines. mutation_testing, dep_audit, intent_verification, large_file_check, migration_reversibility_check, test_expansion, agent_test_mutation, integration_test_gen, review_rubric and their sub-keys were documented and accepted by the schema but never forwarded, so setting them in bernstein.yaml left the gate at its default (#6615). Two consequences for existing seeds: this repository's own bernstein.yaml sets large_file_check: true, which now takes effect, and a quality_gates.intent_verification: that is not a mapping, or an integer key such as timeout_s given a bare true, now raises SeedError instead of being ignored.

Preserve durable MCP tasks on the 2.3 SDK

Upgrade the MCP dependency to the tested 2.3 release line. Keep the existing
2025-11-25 task-create, polling, result, list and cancellation wire contract
backed by Bernstein's authenticated persistent TaskStore, despite removal of
the SDK's experimental task runtime. Ordinary SDK2 clients retain normal tool
calls. This does not implement or advertise the redesigned Tasks extension.

The compatibility adapter uses the SDK's provisional middleware API, so the
dependency is bounded below 2.4 until another compatibility review. Task
cancellation now supplies the existing HTTP endpoint's required JSON body.

The custom HTTP transport retains its four implemented handshake-era protocol
revisions. SDK2 clients in automatic mode negotiate the existing 2025-11-25
wire format; explicitly requested modern HTTP revisions are rejected until
that transport implements their response contract.

Update the Kubernetes client to 37.0.1

Update the Kubernetes extra and its required dependencies. Scope the dependency
review exception for typing-extensions 4.16.0 to its published PSF-2.0 license
metadata, which is already accepted by the project (#6621).

Workflow completion and authority characterization

Adds focused local-process characterization for workflow completion and authority.
Strict expected failures record the proposed completion gaps without changing
runtime behavior; positive controls and process cleanup remain blocking (#6639).

Workflow dispatch crash and restart characterization

Adds focused local-process characterization for workflow dispatch crash and restart.
Strict expected failures record the proposed completion gaps without changing
runtime behavior; positive controls and process cleanup remain blocking (#6639).

dependabot-auto-merge.yml no longer tries to arm auto-merge on a Draft update, which the API refused and which failed the lane; a Draft finishes without calling gh, the new ready_for_review trigger brings it back once it is marked ready, and a Draft state the event did not report as true/false fails closed (#6758).

Security

  • Required maintainer reviews on protected automation and workflow PRs now
    exclude known writers, preserving the existing proven unchanged-diff
    exception and normal governance gates. The same exclusion applies to the
    additional maintainer review for large ordinary changes.

Ship the React Query 5.105.3 client in the dashboard

Update React Query and its core package to 5.105.3, rebuild the bundled
dashboard, and refresh its documented screen captures and bundle binding
together (#6941).

Pi agents stream progress to the runner log

Pi now runs with --print --mode json and receives EOF on stdin, so it
streams task events to the runner log and exits after the prompt instead
of opening its interactive UI or waiting on an inherited stdin pipe.
Log growth supplies the existing liveness signal, preventing heartbeat
escalation from treating an actively streaming agent as idle before its
first artifact (#6943).

Generated adapters match the current CLIAdapter contract

Adapters created from the cookiecutter template now construct successfully
and accept the current spawn arguments. They launch asynchronously and return
the live process handle, PID and log path, with timeout watchdog support.
Generated tests cover the current contract instead of removed result fields
(#6976).

Security

  • Extend the gateway's opt-in initialization capability observation to fingerprint
    complete parsed tool definitions and record per-tool changed fields. Preserve
    legacy name-only audit digests and migrate snapshots without inventing a prior
    definition baseline. This observation does not enforce call holds or approvals
    (refs #6984).

  • Add an explicitly configured persistent operator controller for new TaskStore claims, with fail-closed state validation and serialized halt/resume. This first slice does not stop running workers or implement installation-wide halt.

Type remote status and cancellation observations explicitly

RuntimeBridge now offers optional typed remote status and cancellation
capabilities that refuse unsupported calls without invoking legacy methods.
Frozen DTOs distinguish unknown and waiting states from observed terminal
lifecycle, and unconfirmed cancellation from confirmed cancellation. This
partial component preserves existing bridges; runtime integration and the
observed-event journal remain separate work (#6992).

Added

  • Support an optional * adapter allow-list for roles without a named policy.
    An empty default denies every adapter; existing named-role behavior and
    policies without a default retain their semantics (#7019).

Fixed

  • Apply ordinary contribution review requirements to maintainer-opened pull
    requests carrying commits authored outside the current roster. Co-author
    credit alone preserves the existing rule (#7020).

Preserve offline wheelhouse signing with cosign 3.1.4

Probe offline compatibility flags by parsing them with --help, so flags
hidden from the displayed help still apply. The signing script and real
airgap tests retain the existing detached signatures and offline behavior
when upgrading cosign to 3.1.4 (#7044).

Ship the React 19.3 runtime with the matching browser bundle

Update React and React DOM to 19.3.1 and their type packages to 19.3.0.
Rebuild the bundled dashboard and recapture its documented screens so the
wheel ships the dependency versions recorded in the lockfile (#7052).

  • Use the registered adapter key for admission checks when dispatch falls back to the run pin. Pinned Pi now looks up the existing pi contract. Admission still requires passing evidence and a matching receipt.

Require both trufflehog dependency updates before Renovate opens their grouped branch, and check image extraction and group configuration alongside the existing scanner version gate.

Add an optional bounded render-probe capability to the pinned Playwright driver, with engine-computed styles, checked environment signals, and explicit unavailable results for missing evidence or incomplete captures.

Render comparisons retain unstable properties from both receipts

Render receipt comparisons retain each side's unstable properties, including
unchanged values and incomparable pairs. Existing environment checks and
comparison verdicts are unchanged (#7088).

Show unreadable runs in conduct period summaries (#7093)

Add UnreadableRun and UnreadableRunReason (absent, unverifiable,
torn) to ConductAggregate and aggregate_conduct so supplied run IDs whose
journals could not be read appear in deterministic order alongside valid runs
with unverified coverage, keeping the period coverage unverified.

Conduct aggregates can compare caller-supplied recorded activity scopes against
an effective period scope using the existing authority rules. The read-only
section names divergent axes with source run/activity/event ranges and shows
missing evidence as unavailable or partial.

Goal-drift benchmarks can report recorded intent calibration policy/version,
sample count and existing quality metrics. Receipt replay rederives the report
without fitting a model or changing hard-check results; missing or ambiguous
calibration is explicitly unavailable.

Document the ArtifactSink storage contract (#7169)

Add a short protocol reference for ArtifactSink to
docs/operations/artifacts.md, covering logical key naming, current async
method signatures (write, read, list, delete, exists, stat,
close), lexicographically sorted list() results, and the durable=True
write guarantee.

Cleanup force and inspection are documented separately

Explain selected worktree and branch removal, the additional branch pruning
permitted by bernstein cleanup --force, and best-effort preservation. Show
read-only worktree inspection before cleanup (#7301).

Clarify product and workflow documentation (#7308, #7311, #7312, #7314, #7316, #7318)

The visible homepage, evaluation guide, adapter scope and present-day blog
footers describe Bernstein as an open-source governance layer for AI agents.
The historical ACP entry names the integration it shipped. The workflow
manifest reference distinguishes successful agent dispatch from verified
completion and from successful command execution.

Preserve goal substitution in workflow resume examples

The workflow guide passes the original goal in both resume examples, including
the explicit-manifest form, and lists the saved manifest snapshot fields.
It explains that resume defaults to an empty goal and does not restore it from
the snapshot (#7317).

Clarify README dashboard connections to an existing task server, identify administrative MCP tools and returned run identifiers, and make the name analogy specific to orchestration. Keep all translated README sections aligned.

Artifact-mode completion works on Windows. Its signing identity's Agent Card was written to .sdd/agents/agent:artifact-completion/, and : is not a legal Windows path component, so every artifact-mode task failed there with NotADirectoryError. The card now lives under the SHA-256 of the agent id, the encoding datasource receipt cards already use; the lineage gate matches cards by content, so verification is unchanged. A directory at a task's declared output path now reports "wrote no output" on Windows too, instead of a raw PermissionError.

Contracts for the cline, autohand, openhands, cody, composio and ralphex adapters

Admission refuses an adapter that has no contract. cline, autohand,
openhands, cody, composio and ralphex now have one in tests/contract/contracts/, pinning the flags
each adapter always passes: cline --yolo; autohand --unrestricted --prompt; openhands --headless --override-with-envs --task; cody chat --message;
ao spawn --prompt; ralphex --no-color --task-model. Each was
checked with bernstein adapters contract-check against the current release
of its CLI (cline 3.0.72, autohand 0.9.8, OpenHands CLI 1.16.0, cody 5.5.26,
@aoagents/ao, ralphex v1.7.0).

Count explicit cumulative cost increases accurately

CostTracker.record_cumulative now charges the increase in an explicit
cumulative cost, including corrections with unchanged token counts. Duplicate
or lower reports do not add spend or refund it. Reports without an explicit
cost price each new token once and count toward the next explicit total for
that agent, task and model. New tokens appear in usage reports even when
their monetary delta is zero. Recorded spend drives the existing budget stop.

The persisted format is unchanged. For a key restored from a snapshot, its
first explicit report keeps the previous token-based proration or duplicate
suppression and seeds the monetary baseline for later reports. Estimates
already recorded since the restart count toward that first explicit amount. Exact cost
corrections spanning a restart cannot be recovered from those snapshots.

Fail detached SSH tasks on a nonzero remote exit

Detached SSH execution now records task.failed with the remote exit code
when a task command fails, instead of recording task.completed. The SSH
receipt retains the exit code, and later tasks stay pending. Both the supervisor
and the SSH goal helper close the run as failed. Restarting the supervisor
restores a failed result if it was interrupted after the journal closed.
Concurrent finalizers reuse the first journal outcome and its terminal receipt.
The known limitations also clarify that the
detached local path without an SSH sidecar records task lifecycle events but
does not yet execute the submitted task.

Update reviewdog/action-actionlint to v1.83.0

The Workflow lint job now uses reviewdog/action-actionlint v1.83.0. It keeps the existing reviewdog reporter, error threshold and ShellCheck setting.

Update peter-evans/create-pull-request to v8.1.2

The topology healer, weekly and continuous coverage ratchets, observability snapshot and major/minor release bump use create-pull-request v8.1.2. Their branch names, token selection, inputs and permissions stay as configured. This action uses Node.js 24 on the existing Ubuntu runners.

Update crate-ci/typos to v1.52.0

The spelling job and deterministic auto-heal installation use the same typos v1.52.0 commit. The spelling scan keeps its default read-only mode, and auto-heal retains its strategy condition and explicit write_changes: false input.

A failed workflow node now stops every later layer

bernstein workflow run (and WorkflowRunner.run) used to keep
dispatching nodes in later layers after a node failed, whenever those
nodes depended only on branches that had passed. Those nodes ran their
commands or agents, wrote success checkpoints, and appeared twice in the
result. A failure now skips every later layer once, without dispatching
anything, matching bernstein workflow resume.

Known limitations

  • #4912 (partial): CLI policy loading remains a separate integration task. Injected registries refuse unbound bridge, in-process, container/sandbox and crash-resume paths.
  • Regular fixes ship on 4.0.x. The only planned 3.24.x backport is 3.24.2.

What's Changed

New Contributors

Full Changelog: v3.24.1...v4.0.0