Feature Matrix

August 29, 2026 · View on GitHub

The exhaustive capability index the README links to as its full feature matrix. Every row is verified against src/bernstein/ on the current main.

The "Docs status" column reflects whether a page-level reference exists (Full) or whether the capability is documented in source / module docstrings only (Brief). Full rows link to their reference page.

Rows still marked Brief are deliberate: the capability is reachable from the surface named in its Notes column, but is documented in source and module docstrings rather than on its own reference page. Every row names a surface an operator can invoke. tests/unit/test_cli_server_route_parity.py additionally holds the CLI's server calls to the registered route table.

The "Maturity" column says how far a capability has been proven, which is a different question from whether a page documents it:

MaturityMeaning
5End-to-end evidence, no open correctness issues, docs current
4Tested and documented, minor gaps
3Works, known gaps
2A first-run path is broken
1Stub

The score describes the subsystem behind the row, so rows served by the same code usually carry the same number. bernstein audit and the audit-chain capability rows move together; so do a CLI verb and the capability it drives.

tests/unit/test_feature_matrix_drift.py holds the CLI-commands section to the commands the code registers: a new command with no row here fails that test, and a row naming a command the CLI no longer registers fails it too.


Core orchestration

CapabilityDocs statusMaturityNotes
Goal-based run (-g)Full3Main entry flow
Seed-file run (bernstein.yaml)Full3Auto-discovery supported
Plan-file execution (stages/steps)Full3bernstein run plan.yaml
Retry + escalation plumbingFull3In task lifecycle, with configurable retries
Completion verification (janitor + signals)Full3API + getting started coverage
Process-aware stop/drainFull3Graceful and force stop, drain mode
Multi-cell orchestrationFull3Implemented in multi_cell.py
Fast-path executionFull3Trivial tasks skip the LLM agent entirely (fast_path.py)
Plan mode (human approval)Full3--plan-only, --from-plan, approval routes
Headless modeFull3--headless for CI/overnight
Dry-run modeFull3--dry-run previews the plan without spawning
Typed activity boundaryFull3One hash-in/hash-out contract for coding, research, browser, data, and ops activities, verified by bernstein activity verify <run> (core/orchestration/activity.py). Browser is the one non-coding modality with a CLI dispatch verb (bernstein activity browser run); the rest are Python-API only. Dispatching a non-coding modality from a seed, plan, or backlog file is not wired yet; see Reachability today
Missions (multi-phase goals)Full3Phases run under isolated budget envelopes; a halted phase seals a halt receipt and leaves runnable siblings active (core/orchestration/missions.py)
Durable task suspend/resumeFull3A waiting task parks with an attested receipt that frees its seat, sandbox, and budget; resume reconstructs byte-identically (core/tasks/suspension.py)
Tournament runsFull3Parallel attempts selected by deterministic evaluators; the winner carries a signed selection receipt (core/tournament/)
Fleet steeringFull3Pause, resume, guidance, redirect, and abort each land a signed steering receipt before any effect runs (core/orchestration/steering.py)
Detached run serviceFull3Submit a goal, disconnect, and reattach later against a supervised run service (core/run_service/)
Named resource poolsFull3Lease-backed admission with chain-anchored grant and release receipts (core/admission/)
Spec-to-graph compileFull3bernstein plan compile runs the draft/approve/compile pipeline offline into a gated task graph with a chain-anchored receipt (plan_compile_cmd.py)

State and persistence

CapabilityDocs statusMaturityNotes
File-based state in .sdd/Full3Primary operating model
Metrics/trace persistenceFull3Paths documented, JSONL schema
Lessons/memory persistenceFull4Stored and injected at spawn time
Storage backends (memory/postgres/redis)Full3Config + doctor coverage
Session persistence (fast resume)Full3session.py resumes after stop/restart
Bulletin board (cross-agent messaging)Full3Append-only, used by agents for handoff
Content-addressed artifact storeFull3Content-addressed deduplication for artifacts (core/persistence/cas_store.py)
Cloud artifact sinksFull3Local, S3, GCS, Azure Blob, and R2 sinks behind an async ArtifactSink protocol

Observability

CapabilityDocs statusMaturityNotes
/status and task APIFull3Core API documented
Prometheus /metricsFull3Endpoint is real; Grafana dashboards are user-defined
OTLP telemetry initializationFull3Wiring exists in core/observability/
OTel GenAI span projectionFull3trace project projects a run journal into signed OTel GenAI spans; trace verify-projection recomputes span ids
Live OTLP bridgeFull3The signed span projection streams to any OpenTelemetry collector live or via telemetry export-otel; telemetry verify-span recomputes a span id from its journal entry (core/telemetry/)
Retrospective reporting (retro)Full3CLI coverage present
Cost analysis (cost, history/anomaly hooks)Full4bernstein cost, cost anomaly detection active
Per-agent token progressFull3Tracked in api_usage.py, surfaced in bernstein status
Session analyticsFull3bernstein recap shows session-level stats
Debug bundleFull4bernstein debug collects logs/state/config for triage

Safety and governance

CapabilityDocs statusMaturityNotes
Quality gates (lint, type-check, tests)Full3In the run flow; extended with coverage, benchmark, arch-conformance, and mutation-testing gates
PII scan quality gateFull3Active, auto-installed via log_redact.py
Rule enforcement (.bernstein/rules.yaml)Full3Enforcement behavior documented
Log redaction (PII filter)Full3Active
Lethal-trifecta capability gateFull3Taint-aware egress denial: a chain that unions private data, tainted input, and external comms is refused even when static tags would pass (core/security/capability_matrix.py)
Circuit breakerFull3Halts misbehaving agents, writes SHUTDOWN signal
Token growth monitorFull3Auto-intervention on runaway consumption
Cost anomaly detectionFull3Z-score based, acts via task completion
Agent loop detectionFull3Kills agents in edit-loop cycles
Deadlock detectionFull3Wait-for graph, automatic victim selection
Cross-model verificationFull3A different model reviews completed diffs (opt-in)
Behaviour anomaly detectionFull3Flags agents whose runtime metrics deviate statistically from baseline (core/observability/behavior_anomaly.py)
Context degradation detectorFull3Monitors quality over time, restarts when degraded
Agent trust tiersBrief3bernstein agents trust; tiers accrue from task outcomes in .sdd/trust/ and map to an AgentPermissions profile (core/agents/agent_trust.py)
Volunteer project manifestFull3A project's opt-in policy — OSI licence, acceptance gates, path scope, egress, sandbox floor — loaded and content-addressed by core/volunteer/manifest.py; validate one with bernstein volunteer verify
Volunteer sandbox profileFull3Deny-all-egress containment derived from the manifest and the donor's own limits; refusals are records, not log lines (core/volunteer/sandbox_profile.py)
Volunteer donor budgetsFull3Permanent machine-local task, wall-clock, token, size, and local-model claim limits with atomic restart-safe accounting and signed receipt line items (core/volunteer/budget.py)
Volunteer issue textFull3Untrusted issue title and body normalised into one delimited block before it becomes an agent prompt — HTML comments (closed and unterminated) stripped, invisible and bidirectional characters dropped, NFKC, and a content-derived fence the text cannot forge (core/volunteer/issue_sanitize.py)

Verifiability and provenance

CapabilityDocs statusMaturityNotes
Unified lineage spineFull4When the finalization path emits an external journal seal, artifact provenance and finished-journal identity share that independently committed root (core/lineage/)
Always-on event journalFull3Merkle-chained per-run EventJournal; the head identifies the surviving state, while an external seal establishes complete finished-journal identity (core/replay/journal.py)
HMAC-chained audit logFull4Tamper-evident, daily rotation, cross-process append lock (core/security/audit_chain.py)
Execution WALFull3Hash-chained write-ahead log for crash recovery and determinism fingerprinting
Deterministic replay + forkFull3replay --verify recomputes the journal head; fork --from-step rebuilds state at a journal step into an isolated run (core/replay/)
Audit-chain exportFull4Projects a chain range into COSE_Sign1, in-toto, and DSSE receipts re-verifiable offline with no Bernstein imports (core/security/audit_export.py)
Provenance trust classesFull4Every tool result carries a trust class; effective trust is the minimum over the lineage closure, recomputed offline by audit taint (core/lineage/provenance.py)
Gate adjudication recordsFull3Gate panel decisions are recomputable; gate verify confirms the inputs hash
In-process verification gatesFull3Verification-gate hooks rendered into capable adapters so the check runs inside the agent (adapters/hook_gate_render.py)
Evidence bundlesFull3Sealed verification-evidence bundle, projectable as a tracker comment (core/evidence/bundle.py)
Intent capsulesFull3Approval compiles the goal into a signed capsule; a deterministic drift monitor escalates divergence, verified by intent verify (core/security/intent_capsule.py)
Query receipts (datasources)Full3Read-only SQL results become content-addressed signed receipts; datasource verify --re-execute reports MATCH or DRIFT (core/datasources/)
Compaction receiptsFull3Context and template compaction is recorded as a chain-anchored, reversible receipt (core/tokens/compaction_receipt.py)
Tamper-evident memoryFull4Memory entries are hash-chained with provenance; memory verify/why/forget proves authorship, traces origin, and tombstones (core/memory/chain.py)
Review / autofix / escalation / consent / webhook-node receiptsFull3Signed, journal-anchored receipts verified offline (review-receipt verify, escalation verify, webhook verify)
Result receipt bundlesBrief3A worker submission's patch, gate logs, task ref, and sandbox selection sealed into one DSSE / in-toto envelope; receipt verify recomputes it offline and names the field that diverged (core/security/result_receipt_bundle.py)
Stall escalation receiptsFull3A stalled worker produces a signed escalation receipt embedding the last audit entries and a deterministic recommended action (supervisor escalate)
C2PA content credentialsFull3Artifact lineage projected into signed C2PA credentials (credential emit/verify)
Skill install receiptsFull3Install and usage links recomputable via skill verify / skill provenance
Adapter security-floor receiptsFull3A below-floor adapter spawn is refused with a content-addressed, chain-anchored refusal receipt (adapters/security_floor.py)
Endpoint certificationFull3Local-model workers are conformance-tested and issued a signed certification (core/endpoints/certification.py)
SLA violation receiptsFull3Per-goal SLA contracts evaluated read-only each tick; a breach becomes a signed, offline-verifiable violation receipt (core/planning/sla_store.py)
Signed daily mission digestsFull3Each day's mission progress is a signed digest projecting that day's chain; mission digest verify re-derives it (core/orchestration/mission_digest.py)
Agent run manifestFull3Hashable workflow spec for SOC 2 evidence
RBAC / budget / seat projectionsFull3Access and budget verdicts re-derivable via governance verify
Unified event feedFull3Chain-projected event feed with a typed grammar and per-event receipts (core/events/feed.py)

Identity and delegation

CapabilityDocs statusMaturityNotes
Native subagent delegationFull4The scheduler delegates leaf execution to native subagents; schema-validated results anchor as subagent.delegation journal entries, chain verified via delegation verify (core/agents/subagent_delegation.py)
Attenuated capability tokensFull4Each delegation hop is a signed, scope-attenuating token, so the principal to orchestrator to sub-agent authority chain verifies offline (core/security/capability_tokens.py)
Signed agent cardsFull4AgentIdentityCard signed for A2A federation and served at /.well-known/agent-card.json (core/security/agent_card_signer.py)
A2A node + message receiptsFull4Callable JSON-RPC node; a completed task returns an artifact whose parts carry a lineage receipt verified by a2a verify (core/interop/a2a_card.py)
HTTP message signaturesFull4RFC 9421 Ed25519 signatures on outbound agent-facing requests; keys served as JWKS via identity keydir (core/identity/http_signing.py)
SPIFFE workload identityFull4SPIFFE/SVID workload identity with spiffe id and spiffe verify-binding (core/identity/spiffe/)

Payments and cost governance

CapabilityDocs statusMaturityNotes
Spending mandatesFull3Authorized-spend mandates enforced before a paid action, with signed spend receipts (core/payments/)
Authorized-action mandatesFull3mandate emit/verify/revoke binds, proves, and revokes an authorized-action mandate
Cost-policy receiptsFull4The live dispatch gate records batch, cache, and model policy verdicts as receipts (core/cost/scheduling/receipt.py)
Budget envelopesFull4Per-phase and per-task budget envelopes with rollup by envelope (core/cost/cost_rollup_by_envelope.py)

Ecosystem and integrations

CapabilityDocs statusMaturityNotes
Agent catalog/discoveryFull3bernstein agents sync/list/discover/match/showcase (40+ CLI agent adapters)
Browser / computer-use adapterFull3Adapter family for autonomous browser and computer-use agents; every action is a content-addressed lineage anchor (adapters/computer_use.py)
GitHub App and CI fix flowsFull3bernstein ci fix <url>, github setup
Trigger sourcesFull3github, gitlab, slack, discord, file_watch, webhook, odata, and schedule source adapters (core/trigger_sources/)
OData trigger sourceFull3Polls an OData endpoint into normalized trigger events (core/trigger_sources/odata_poll.py)
Webhook nodeFull3Outbound webhook node with recomputable inbound-event and node hashes (core/trigger_sources/webhook_node.py, webhook verify)
Automation bridgeFull3Signed trigger receipts and chain-anchored status callbacks for external automations (core/trigger_sources/automation_platforms.py)
Plugin hooks (pluggy)Full3SDK docs in CONTRIBUTING.md
Cluster/worker primitivesFull3bernstein worker --server URL, cluster routes documented. A worker registering against a live cluster-enabled server is covered by tests/integration/test_first_run_long_running_surfaces.py, which is the same code path the bernstein worker CLI row rests on.
Multi-repo workspacesFull3workspace: in bernstein.yaml, workspace CLI
MCP server modeFull3bernstein mcp, MCP server in mcp/server.py
MCP tool registryFull3Auto-discovery and per-task config
MCP catalog clientFull3bernstein mcp catalog browse/search/install installable server catalog (core/protocols/mcp_catalog/)
MCP input contractsFull3Schema-validated, deny-by-default MCP tool-call input firewall (mcp/input_validation.py); the enforced schema is also the advertised inputSchema
Stateless MCP anchoringFull3A stateless MCP client can poll a run and verify it offline; calls anchor as mcp.stateless_call journal entries (core/protocols/mcp/stateless_core.py)
Runtime capability cards (MCP)Full3Per-server capability cards for the MCP server (mcp/capability.py)
ACP native bridgeFull3bernstein acp serve --stdio|--http :PORT IDE-native bridge (core/protocols/acp/); see reference/acp-bridge.md
Protocol negotiationFull3protocol_negotiation.py runtime protocol-version handshake
Schema registryFull3schema_registry.py versioned message schemas for protocols
Credential vaultFull3bernstein connect <provider>, bernstein creds list/revoke/test OS-keychain token storage (core/security/vault/)
Autofix CI daemonFull3bernstein autofix start/stop/status/attach watches PRs and dispatches repair runs on CI failure (core/autofix/)
Dev previewFull3bernstein preview start/stop/list/status exposes an agent dev server via tunnel with configurable auth (core/preview/)
Fleet dashboardFull3bernstein fleet [--web HOST:PORT] cross-session multi-instance view (core/fleet/)
Notification sinksFull3bernstein notify test --sink <id> pluggable notification backends (core/notifications/)
PR review responderFull3bernstein review-responder start/status/tick auto-responds to PR review comments (core/review_responder/)
Review pipeline DSLFull3bernstein review --pipeline review.yaml YAML-driven multi-phase review, plus the bounded --fix --until-checks-green contour and its per-pass receipts (core/quality/review_pipeline/)
Plan archivalFull3bernstein plan ls/show list and inspect archived plans (core/planning/lifecycle.py)
Slack integrationFull3Slash commands and events API endpoints
Webhook ingestionFull3POST /webhooks/ for external event routing
Adaptive parallelismFull3Auto-tunes concurrency from observed success rates (core/orchestration/adaptive_parallelism.py)
Warm poolFull3Pre-spawned agent pool to cut spawn latency (core/agents/warm_pool.py)
Pluggable sandbox backendsFull3Worktree, Docker, E2B, and Modal backends behind a SandboxSession protocol
microVM sandbox backendBrief3Isolation tier with content-addressed snapshots for deterministic fork-and-race (--sandbox microvm)
Named sandbox poolsFull3Chain-projected pool manifests with capability and egress ceilings and signed worker enrolment (core/sandbox/pool.py)
Cache policy engineFull3Content-addressed key recipes, drift expiry, and fleet dedup with signed duplicate-of receipts (core/persistence/cache_policy.py)
Sovereign deployment profileFull3Signed residency-posture attestation; posture drift at spawn is a signed refusal (core/security/deployment_profile.py)
Workflow DSLFull3bernstein workflow validate/list/show
Chaos engineeringFull3bernstein chaos agent-kill/file-remove/status/slo
Benchmark suiteFull4bernstein eval run/compare/swe-bench/programbench
Eval harnessFull4bernstein eval run/report/failures
SWE-Bench harnessFull4Verified eval in benchmarks/swe_bench/run.py
Graduation systemFull3Agent promotion stages, routes in routes/graduation.py. The documented REST surface is driven end to end against a real server by tests/integration/test_first_run_graduation_surface.py: policies, record-event, eligibility at the sandbox threshold, and promotion to shadow.
Semantic cachingFull3semantic_cache.py prompt deduplication
Cascade router (intra-Claude tier escalation)Full3Tier escalation within a single provider (core/routing/cascade_router.py)
Cascade fallback manager (cross-adapter failover)Full3Cross-adapter provider failover (core/routing/cascade.py)
Batch routerFull3Task batching for non-urgent work
Prompt cachingFull3SHA-256 system prefix deduplication
Output style customizationBrief3.bernstein/output-styles/*.md, selected by output_style: in bernstein.yaml, injected into the spawn prompt
Installation mismatch detectionFull4Detects adapter/installation gaps
Worker badge identityFull3Process identification in ps/Activity Monitor
Keybinding system (TUI)Full3Configurable TUI keyboard shortcuts. tests/unit/test_first_run_keybinding_overrides.py resolves the documented three layers and then presses the overridden key against a running Textual app, asserting the bound action fires and the pre-override key no longer does.
Diff folding displayBrief3bernstein diff --fold; the TUI agent log also folds long diffs in its historical tail
Word-level diff renderingBrief3bernstein diff --word-diff highlights only the tokens that changed
Contextual tips systemBrief3One cooldown-limited hint after an interactive command; BERNSTEIN_NO_TIPS opts out
Security review commandBrief3bernstein security-review pattern-scans a diff for secrets, injection, and weak crypto
Commit attribution statsFull3Per-agent commit statistics, via bernstein report commits
Away summary generationBrief3bernstein recap --since 6h builds the report from workspace files, no server needed
Plugin trust warningBrief3Trust tier and score per plugin in bernstein plugins (--trust-details for the signal breakdown)
Cumulative progress trackingFull3Progress tracking across runs

CLI commands

CommandDocs statusMaturityNotes
bernstein -g GOALFull3Inline goal
bernstein run plan.yamlFull3Plan file execution
bernstein gc casFull3Mark and sweep unreferenced blobs from the CAS store
bernstein initFull3Workspace setup. The documented first run is covered by tests/integration/test_first_run_documented_path.py, which runs the command from an empty directory and asserts the created artifacts plus the key output lines documented in first-run.md.
bernstein stopFull3Graceful/force stop
bernstein liveFull3TUI dashboard. Readiness is the first rendered frame, identified by the AGENTS and TASKS pane headers; tests/integration/test_first_run_long_running_surfaces.py starts it from an empty workspace, waits for that frame, and asserts a traceback-free exit on SIGINT.
bernstein dashboardFull3Web dashboard
bernstein statusFull3Task summary
bernstein psFull3Process list
bernstein costFull4Spend breakdown
bernstein doctorFull4Pre-flight health check
bernstein recapFull3Post-run summary
bernstein retroFull3Retrospective report
bernstein runs reportFull3Finished runs projected from the work ledger and classified pr-opened / gate-failed / no-changes / infra-error / wedged, each with the evidence line it was read from
bernstein report commits/incident/postmortemBrief3Per-run markdown summaries: commits is per-agent commit attribution (reference); incident correlates a timeline from logs, metrics, and traces; postmortem writes a structured report for a failed run. The group has no reference page of its own; cli-reference.md and bernstein report --help carry it
bernstein trace IDFull3Decision trace
bernstein logsFull3Agent log tail
bernstein diff IDFull3Per-task git diff
bernstein planFull3Task backlog
bernstein plan compileFull3Compile a spec into a gated task graph offline
bernstein replay IDFull3Deterministic replay
bernstein checkpointFull3Session snapshot
bernstein wrap-upFull3End session with summary
bernstein demoFull2Preview. Zero-config demo. A cold first run exits 1 with Task server on port 8055 did not respond within 10.0s and an empty server.log: a first-ever import of the package can exceed the server-readiness budget while Python byte-compiles it. The real budget is 30s (_SERVER_READY_TIMEOUT_S); the 10.0s in the message is a stale hardcoded string — #3905. A second run on a warm cache starts the server in under a second. On a Windows console using a legacy code page the command also aborts with UnicodeEncodeError on the it prints first.
bernstein demo --flask-todoFull2Preview. Flask TODO demo (3 tasks). bernstein quickstart is a deprecated alias, removed in v4.0.0. The run does not complete a first run: all three seeded tasks fail and the task server can die and restart mid-run. The reporting is no longer part of it — the summary counts against the seeded total and the command exits non-zero (#3902).
bernstein agents ...Full3Catalog management
bernstein evolve ...Full2Preview. A clean directory exits before the evolution loop starts because .sdd/ is missing; initialise a Bernstein workspace first.
bernstein ci fixFull3CI autofix
bernstein github setupFull3GitHub App setup
bernstein workerFull3Join cluster as worker. Readiness is the Registered as node <id> line; tests/integration/test_first_run_long_running_surfaces.py registers one against a live cluster-enabled server and separately asserts the non-git-workspace refusal names its reason.
bernstein mcpFull3Run as MCP server
bernstein chaosFull3Fault injection
bernstein auditFull4Cryptographic audit chain. Seventeen subcommands: seal, verify, verify-hmac, verify-gates, verify-multitenant, verify-suspension, show, query, diagnose, receipt, taint, capabilities, ack-tear, export, pack, slice, archive
bernstein verifyFull4Merkle/HMAC verification
bernstein benchmarkFull4Deprecated alias for bernstein eval, removed in v4.0.0
bernstein evalFull4Evaluation harness
bernstein workspaceFull3Multi-repo workspace
bernstein configFull3Configuration management
bernstein quarantineBrief3Cross-run task quarantine read from .sdd/runtime/quarantine.json (list, clear)
bernstein cacheFull3Response cache management
bernstein test-adapterFull3Adapter smoke test
bernstein add-taskFull3Inject task via CLI
bernstein cancelFull3Cancel task
bernstein review/approve/reject/pendingFull3Review workflow
bernstein syncFull3Sync backlog with server
bernstein manifestFull3Run manifest inspection
bernstein gatewayFull3MCP gateway proxy
bernstein workflowFull3Workflow DSL
bernstein watchFull3Directory file watcher
bernstein listenFull2Preview. Voice commands are experimental. A base install exits with the pip install 'bernstein[voice]' hint; a usable first run also requires microphone/audio support and downloads the selected Whisper model on first use.
bernstein completionsFull3Shell completion scripts
bernstein selfFull3Provenance-verified update lifecycle: signed feed, chain-anchored advisory, pre-install wheel verification, pin, receipted rollback
bernstein self-updateFull3Compatibility alias for bernstein self
bernstein pluginsFull3List active plugins
bernstein install-hooksFull3Install git hooks
bernstein debugFull4Generate debug bundle for triage
bernstein acp serveFull3ACP bridge (--stdio or --http :PORT)
bernstein autofix ...Full3CI autofix daemon (start/stop/status/attach)
bernstein connectFull3Credential vault setup for a provider
bernstein creds ...Full3Credential management (list/revoke/test)
bernstein preview ...Full3Dev server preview (start/stop/list/status)
bernstein fleetFull3Fleet dashboard (optionally --web HOST:PORT)
bernstein fleet steerFull3Mid-run steering: pause/resume/guidance/redirect/abort
bernstein mcp catalog ...Full3MCP catalog browser (browse/search/install)
bernstein notify testFull3Notification sink smoke test
bernstein plan ls/showFull3List and inspect archived plans
bernstein review-responder ...Full3PR review responder (start/status/tick)
bernstein review --pipelineFull3Review with YAML pipeline DSL
bernstein fork --run --from-stepFull3Fork a run at a journal step into a new isolated run
bernstein gate verifyFull3Recompute a gate panel's inputs hash and confirm the adjudication
bernstein mandate emit/verify/revokeFull3Bind, prove, and revoke authorized-action mandates
bernstein payment-mandate issue/show/spendFull3Issue and spend against authorized-spend mandates with signed receipts
bernstein governance verifyFull3Recompute access and budget verdicts for a run
bernstein webhook verifyFull3Recompute inbound-event and outbound webhook-node hashes
bernstein review-receipt emit/verifyFull3Bind and offline-verify PR review receipts (issue + plan + tool calls + diff)
bernstein receipt create/verifyFull3Sign a result receipt bundle from a JSON spec and verify one offline, naming the exact field that diverged
bernstein escalation show/verifyFull3Project and reconstruct escalation receipts from the journal
bernstein supervisor status/escalateFull3Supervise stalled workers and seal stall escalation receipts
bernstein delegation verifyFull4Reconstruct and verify a run's delegation chain
bernstein credential emit/verifyFull3Project an artifact's lineage into a signed C2PA credential and verify it
bernstein skills provenance/verifyFull3Recompute a skill's install receipt and usage-provenance graph
bernstein schedule verify/audit, schedule show --atFull3Replay recorded fires, chain-check fire receipts, project a schedule at a time
bernstein sla add/list/show/verify/reportFull3Attach per-goal SLA contracts; a breach is an offline-verifiable violation receipt
bernstein trace project/verify-projectionFull3Project a run journal into signed OTel GenAI spans and verify the projection
bernstein telemetry export-otel/verify-spanFull3Stream the span projection to an OTLP collector and verify a single span offline
bernstein thread verifyFull3Prove a streamed TUI thread equals its executed journal
bernstein memory verify/why/forgetFull4Prove authorship, trace origin, and tombstone a memory entry
bernstein replay --verify/--from-stepFull3Recompute the journal head or rebuild state to a step
bernstein lineage verify/walk/chain/replay/exportFull4Per-artifact lineage spine: recompute a run's Merkle chain and HMAC tags, walk an artifact back to its producer, replay the spine in append order, and export a run's chain as a regulator artefact (tracker-audit, forks, conflicts, resolve, merge, reindex, gate, and v2 complete the group)
bernstein intent show/verifyFull3Project and recompute an intent capsule's conformance offline
bernstein a2a verify/publishFull3Verify an A2A message receipt and publish the agent card
bernstein activity verifyFull3Verify a typed activity's replay hashes (activity browser run for browser checks)
bernstein datasource register/query/verifyFull3Register read-only SQL datasources and verify content-addressed query receipts
bernstein evidence show/verifyFull3Project and verify a sealed evidence bundle
bernstein events query/verifyFull3Query the unified event feed and verify its chain projection
bernstein endpoints certify/verifyFull3Conformance-certify a local-model endpoint and verify its certification
bernstein ledger verify/anchor/fetchFull3Verify, anchor, and fetch work-ledger segments
bernstein mission define/status/verifyFull3Define multi-phase missions and verify mission status (mission digest verify for digests)
bernstein tournament show/verifyFull3Inspect a tournament run and verify its selection receipt
bernstein spiffe id/verify-bindingFull4Print the SPIFFE id and verify a workload-identity binding
bernstein spec check/auto-fixFull3Evaluate and auto-fix a spec against the quality checklist
bernstein run-service submit/attach/statusFull3Submit a detached run, then reattach to it later
bernstein compaction logFull3Inspect chain-anchored compaction receipts
bernstein identity keydir/decode/verifyFull4Print the JWKS key directory and decode/verify install identity
bernstein pool register/list/show/verifyFull3Manage lease-backed named resource pools
bernstein volunteer verifyFull3Validate a project's .bernstein/volunteer.json and print the manifest digest a receipt binds to
bernstein volunteer budgetFull3Set or inspect persistent donor limits and completed/in-flight usage

Cloud / Cloudflare

How a row graduates: A row graduates out of Preview when its maturity score increases to ≥ 3 (or when a first-run smoke test lands and the marker is intentionally removed).

CapabilityDocs statusMaturityNotes
Workers RuntimeBridgeFull2Preview. Hosted API (api.bernstein.run) is unreachable; Workers run against your own Cloudflare account. bridges/cloudflare.py agents on Workers + Durable Objects
Workflow Bridge (durable execution)Full2Preview. Hosted API (api.bernstein.run) is unreachable; Workflows run against your own Cloudflare account. bridges/cloudflare_workflow.py auto-retry, approval gates
Browser Rendering BridgeFull2Preview. Hosted API (api.bernstein.run) is unreachable. bridges/browser_rendering.py screenshots, scraping, PDFs
R2 Workspace SyncFull2Preview. Hosted API (api.bernstein.run) is unreachable; R2 sync runs against your own Cloudflare account. bridges/r2_sync.py content-addressed delta sync
Workers AI Provider (free LLMs)Full2Preview. Hosted API (api.bernstein.run) is unreachable. core/routing/cloudflare_ai.py Llama, Mistral, Gemma, Qwen
MCP Remote TransportFull2Preview. Hosted API (api.bernstein.run) is unreachable. mcp/remote_transport.py streamable HTTP for remote MCP
Cloud CLI (bernstein cloud)Full2Preview. bernstein cloud login/run/status/runs/cost report the service unreachable and exit non-zero. cli/commands/cloud_cmd.py login, run, status, cost, deploy
Codex-on-Cloudflare AdapterBrief1Preview. Targets a REST API that does not yet exist and refuses fast. adapters/codex_cloudflare.py experimental