AgentWorkflows v0.10.0

Docs: https://ramazankara.github.io/agentworkflows/

A platform-wide overhaul: approval escalation, a gateway that adds a third of its former
per-request overhead, a faster and sturdier console, TypeScript SDK parity with Python,
BM25 retrieval, faster run listings and insights, and a release pipeline that checks every
version reference before it publishes. Images, charts and SDKs are version 0.10.0.

Behaviour changes to know about

• A workflow attempt that no provider processed (connect error, 429, 503 or an adapter
  refusal) now gives its run budget back; ambiguous failures keep the charge.
• The gateway gzips JSON, JavaScript and CSS responses over 1 KiB for clients that accept
  it. Event streams, images and fonts are untouched.
• rag-ingest --write deletes the stale chunks of the sources it ingests; pass
  --no-prune for the old upsert-only behaviour.
• The TypeScript client retries state-changing calls when the gateway provably did not act
  (refused connection, 429 or 503) and rejects non-http(s) base URLs.

Gateway

• Lower overhead per request, measured against a bare FastAPI app with a fake runtime:
  1879 to 559 µs by default, 2474 to 730 µs on the governed path and 3090 to 986 µs when
  streaming. A team chat request makes 9 Redis round trips instead of 14: the rate limiter
  is one atomic script, budget settlement and provider usage share one, and team settings
  are read once per request.
• A client that disconnected before the response body no longer leaks the team request
  lease, which could leave the team impossible to erase.
• Team audit-view rows are written off the event loop in chain order, so a slow Redis or
  Postgres no longer stalls every request, and audit chain heads are saved from a
  consistent snapshot, which prevents false broken-link findings after a restart.
• Stream receipts record a client disconnect as 499 and a mid-stream error as 500 instead
  of 200. A provider-key store outage returns a retryable 503, lone UTF-16 surrogates are a
  4xx, and malformed headers get the standard error envelope.
• The in-memory rate-limit and budget maps drop finished windows, an expired JWKS cache
  sends one refresh to the identity provider, and batch file downloads stream.
• Console assets are cached as immutable and gzipped, so the console's script reaches the
  browser at under a third of its size and is not downloaded again on a repeat visit.

Approval escalation

• Team admins can escalate a workflow's approvals that wait too long. After after_seconds
  the review notifies the team's alert channels (or channels chosen for the escalation) and
  opens to more reviewers: every approver, or named reviewer identities. The escalation and
  every vote it admits are receipted. Nothing changes unless escalation is configured, and
  in-flight runs on existing workers escalate too.
• API: GET /v1/team/approval-escalations and PUT/DELETE /v1/team/workflows/{name}/approval-escalation. Run views include approval_escalation
  and the caller's approval_access.
• Python SDK and CLI: approval_escalations(), set_approval_escalation(),
  remove_approval_escalation(), review_queue(), workflows escalation list|set|remove
  and runs pending. TypeScript: approvalEscalations(), setApprovalEscalation(),
  removeApprovalEscalation() and reviewQueue().
• Console: Approvals opens on "Ready for you", the drafts you can decide now; "All waiting"
  shows the rest and when escalation opens a draft to you. Runs show when they escalate and
  an Escalated mark, timelines and the audit log name escalation receipts and votes, and
  Team settings has an escalation editor that flags an escalation the approval expiry would
  overtake.

Console

• Each page loads on demand: the shell script is 246 kB instead of 374 kB (78 kB instead of
  110 kB gzipped), and Inter ships only its Latin and Latin Extended files (133 kB instead of
  218 kB). A 10-page admin session makes 24 API calls instead of 34.
• A failed refresh no longer blanks a run page or stops its auto-refresh. Polling pauses in
  hidden tabs and waits for each answer. Pages that fail to load offer Try again instead of
  showing Loading forever, and a page whose code fails to load offers a reload.
• Saving alert rules, retention or an escalation and then team settings on the same page no
  longer fails with a stale-revision conflict, nor does registering a workflow right after a
  template install.
• Rejecting a draft asks for confirmation, focus moves to the next draft after a decision,
  and leaving with unsaved team settings, an uncopied API key or an uncopied invitation link
  asks first. Workflow result links open only http(s) addresses.
• Run filters and the Insights period live in the address, Insights and Costs tables sort by
  column, Members & keys has a search box, and the audit log pages back to newer events.
• Clearer error messages, including a separate one for gateway errors.

Workflows and Python SDK

• Finished runs are served from a snapshot. Listing 100 runs takes 78 ms instead of 688 ms,
  insights over 120 runs 111 ms instead of 235 ms, and the 30-second monitor scales with live
  runs instead of retained ones (107 ms instead of 1365 ms for 120 runs). A run timeline
  reads its step content in one batch.
• Fixed: scheduled triggers waiting on a long review hit Temporal's history limit after about
  38 hours and copied run content into schedule history. Polling now backs off from 30
  seconds to 5 minutes, continues as new and stores only the run ID and status. Both SDKs.
• Signed webhook replay records expire after 8 days instead of being kept forever.
• Cancelling a finished run returns 409 workflow_not_cancellable instead of a 404, and a
  retried vote no longer writes a duplicate approval receipt.
• agentworkflows --help starts in 106 ms instead of 274 ms.
• The Python client no longer repeats run, cancel, retry, delete, pause or install calls
  after an ambiguous 5xx or timeout. Container steps reject workspace names that are not
  Kubernetes names, and AGENTWORKFLOWS_HEALTH_ADDRESS sets or turns off the worker's
  health endpoint.

TypeScript SDK

• GatewayClient matches the Python client: chat and streaming, completions, embeddings,
  moderations, batch, Messages, Files, Batch and Responses APIs, receipts, team, usage and
  readiness, with sandboxId, headers, retryBackoffMs, signal, fetch and per-call
  withOptions().
• WorkflowGateway.agent() runs OpenAI, Anthropic and other framework agents as governed
  steps, and WorkflowGateway.container() runs container steps with the Python SDK's
  workspace checks.
• Typed GatewayTimeoutError, transport error codes that are safe to log, and state-changing
  calls that retry only when the gateway provably did not act.
• runWorker serves the chart's health probes on request. Subpath types work for consumers
  on moduleResolution: node, and every pack checks CommonJS and ESM imports.

RAG service

• Lexical retrieval ranks with BM25 in any script. On the docs corpus, recall@3 went from
  0.17 to 0.96 and a query takes 0.3 ms instead of 12 ms. Hybrid scoring ignores stopwords,
  and the grounded context includes every retrieved document.
• Ingest batches embeddings and Qdrant upserts (5x faster on 627 chunks, with request bodies
  capped near 1 MB), stamps chunks with their document, and --write prunes the stale chunks
  of the sources it ingests (--no-prune keeps them). scripts/rag-ingest.py now runs this
  same CLI.
• Malformed vector-store or embedding responses return 503 with Retry-After, a lone
  surrogate in a query returns 422, a bad knowledge file or failing bootstrap no longer takes
  the service down, non-string tenant claims are refused, and settings errors name their
  environment variables.
• The audit chain head is persisted from a consistent snapshot, so audit-verify no longer
  falsely flags a link across restarts.

Release, CI and deployment

• make release-prep RELEASE_VERSION=X.Y.Z moves every version reference at once and make release-check proves they agree. The release workflow checks before it publishes, builds
  the three images in parallel, and attaches SBOM and provenance attestations.
• Fixed: the umbrella chart's lock still pinned 0.9.0, and published charts lost their
  values.yaml comments.
• CI's single job now also covers the TypeScript SDK and its package check, Ruff
  formatting, and mypy for the gateway, RAG service and Python SDK. make static-checks
  runs the Helm, Compose, contract and policy checks locally in about a minute.
• Multi-stage images are smaller (worker 334 to 307 MB) and rebuild in seconds when only the
  revision changes. Bundled PostgreSQL has startup and liveness probes, Temporal jobs have
  resource requests, and Compose services report healthy faster.