Harness release candidate

This is an additive source snapshot on public OpenWorker main. It is a branch for integration testing, not a release announcement or a claim of stable APIs.

Scope

Rust runtime, Rust/TypeScript SDKs, native Tools SDK, SQLite/PostgreSQL storage, server reference, artifacts, telemetry and deployment references.
Arcade Studio and the minimal desktop starters remain available.
Lantern/Luma application code, custom app runtimes and obsolete app-specific partner-kit packaging are omitted. Notification/device/computer tool crates, core approval/MCP tests and independent stdio framing tests remain.
Private partner notes, credentials, raw deployment logs and generated state remain excluded. Public implementation ledgers and sanitized qualification evidence accompany the service integration. AWS scripts require an explicit target account/role and budget notification recipient.
Existing Python Desktop source is unchanged. Runtime fixtures now live inside harness/tests/contracts/; the Rust workspace does not need an external fixture directory just to compile.

Layout migration

The Rust workspace is harness/; public guides are docs/harness/. Update local path dependencies, executable paths, build caches and scripts that referenced an older directory layout. Rust crate names and @openworker/harness remain the same. Generated partner workspaces contain path dependencies: regenerate them against the new workspace instead of editing installed binaries.

Synthetic examples use examples/assistant and example_echo; reusable native tools use notifications.* and device.*. These example names are not aliases for private application tools. Existing installed applications remain on their pinned revision until their owner deliberately updates and requalifies them.

Verification for this snapshot

Checks repeated during the 2026-10-26 source-layout migration:

Rust formatting and strict workspace Clippy; SDK/custom-host compilation.
Linux runtime unit/integration suites, including native tools, Anthropic fixtures, compaction, approvals, SQLite teams, model connections, MCP and resource grants.
Mac SDK desktop/provider-auth tests and the retained native tool crates.
Deterministic server, filesystem-artifact and telemetry tests. The opt-in S3 integration test was not run because no object-store fixture was configured.
TypeScript configuration/transport tests and real-host provider-auth/OpenRouter streaming fixtures. These fixtures use local HTTP servers, not live provider keys.
Independent MCP acceptance: qualified tool selection, SSE discovery, approvals, restart, Q&A, denial, redirects, cancellation and external-server ownership.
Arcade: 27 tests and the production frontend build with Node.js 24. A freshly generated Electron starter also installed dependencies and built its frontend.
Starter/custom-host generator tests, retained raw-stdio tests, deployment-script tests and Task Desk/console tests. One startup-cancellation test was skipped because its opt-in stack fixture was not configured.
Release-layout checks: workspace members, removal of legacy apps, relative documentation links and source-checkout starter commands. Gitleaks found no secrets in the staged snapshot. The branch starts from public main without importing private commit history.
Linux SDK/PostgreSQL matrix: 123 passed, zero failures, against an isolated PostgreSQL 16.4 Docker fixture. Includes SQLite/PostgreSQL embedded/subprocess team parity, dynamic agents, auth rotation, ownership fencing, durable controls, queue ordering, authorization, worker death, database loss and draining. The opt-in load benchmark and mixed-version upgrade drill were explicitly excluded; the latter requires a separately qualified baseline worker binary.
The SQLite embedded/subprocess agent regression also passed in a separate, serialized run with the full workspace test-feature configuration.

The HTML QuickStart was checked at the source/link level. A fresh browser render was blocked by the local file-URL policy. No fresh GUI, signed installer, Windows VM, live-model, AWS or OpenShell cluster qualification is claimed by these checks. Generator dependency installation used npm 10 with an engine warning; use the documented npm 11+ for supported starter development.

Service integration baseline — October 30, 2026

openworker-harness-sdk now includes feat/openworker-service-m1 through eee4391f. The integration is a fast-forward: the runtime/SDK/service source is identical to the qualified feature-branch source, with this checklist and its ledger updated afterward. This remains an integration branch, not a public-main merge or release announcement.

Authenticated service clients, RuntimeProfiles, OpenShell provisioning, management UI, ows deployment commands, durable artifacts and observability are included.
Observability M1–M5 qualification covers PostgreSQL recording, model accounting, authorized run search, management charts/timelines, collector export and Langfuse reconciliation. See the observability ledger.
A real Claude Sonnet 5 agent ran file/shell tools inside local Kubernetes / OpenShell. All 22 run-scoped records and 10,560 input / 222 output tokens matched Langfuse; the management UI and Langfuse UI were visually verified. Records survived replacement of all four Langfuse data-service pods with retained PVCs.
The catalog-worker instrumentation fix passed the real PostgreSQL/OpenShell service suite, 13 server-reference tests and eight Helm tests. Earlier outage trials retain their gaps; final qualification used a fresh explicit turn.
The next cloud gate is a fresh AWS installation of this consolidated baseline, including real agent execution, recovery, telemetry and expiry cleanup. Earlier AWS qualifications do not certify this newer baseline. GCP and Azure require their own deployment qualification.

Evidence: live dual-UI result, persistent local setup, service progress.

Canonical integration branch consolidation — October 30, 2026

Fast-forwarded openworker-harness-sdk to the verified GCP feature work through e4f96969, preserving all four commits after the earlier 6e1bc173 baseline. There was no branch divergence or source conflict. This is the canonical public feature-development and partner-integration branch; tested feature milestones should return here regularly instead of accumulating only on working branches.

GCP identity bootstrap/direct preflight, private Terraform graph export, local lifecycle journals and the release-size baseline are now included.
The exact merged CLI source passed 75 tests and real pinned Terraform/provider schema validation before integration. The fast-forward does not change that source; subsequent consolidation changes are documentation only.
At this consolidation checkpoint, GCP remote provisioning and independent cleanup were unimplemented gates. The local lifecycle milestone below supersedes that implementation status without claiming a qualified GCP deployment.
The Python desktop harness is unchanged. This consolidation does not merge into public main, publish packages or announce a stable SDK release.

See the GCP ledger and binary footprint baseline. Partner builds should pin a revision and keep SDK/runtime versions aligned.

GCP lifecycle recovery and independent cleanup — October 30, 2026

Fast-forwarded openworker-harness-sdk to verified milestone c918b486 from feat/ows-gcp-lifecycle. Changes are confined to the management CLI, cleanup container and documentation. The runtime, Rust SDK and TypeScript SDK sources are unchanged from the previous integration checkpoint.

95 CLI tests pass, including loopback transport, lost responses, checkpoint compare-and-swap, ownership fences and scheduler pause reconciliation.
Pinned Terraform validates both control foundation and runner graphs with zero errors/warnings. A Linux AMD64 cleanup image builds and passes network-disabled, read-only, unprivileged container smoke checks.
Production workload creation remains closed in code. No new cloud mutation, IAM grant or paid resource was created during this milestone. Full preflight, source/control installation and live independent cleanup remain to be qualified.

See the lifecycle guide, progress ledger and sanitized evidence.

GCP control-artifact installation — October 30, 2026

The verified installation milestone adds private locked journals, read-only resolved preflight, loss-safe generation-pinned source/control/checkpoint uploads, actual receipt-based cleanup packages and existing job/heartbeat verification. 115 CLI tests pass; both control graphs validate with pinned Terraform/Google schemas. No GCP Clippy diagnostics. Runtime/Rust SDK/TypeScript SDK sources are unchanged.

This is not a live GCP deployment qualification. Foundation/job Terraform apply, scoped effective authority, fresh cost/prior-spend evidence, private installer access and OpenShell/live-cleanup qualification remain outstanding. No new cloud mutation, IAM grant, API enablement or paid resource in this milestone.

See installation guide, ledger and evidence.

Gates before announcement

Several days of Arcade/partner dogfooding, including real provider streaming, auth expiry, pending approvals, cancellation and restart recovery.
Clean-checkout builds and packaged app acceptance on target Windows/Mac OS and CPU architectures; code signing/notarization and runtime prerequisites.
Explicit Linux/PostgreSQL/OpenShell deployment qualification for the selected versions. Local fixture tests do not certify AWS/Azure or every cluster policy.
Real local-model/tool compatibility, image consent and failure UX.
Dependency/asset license review and release version/package publication decisions.
Review public API/source compatibility before tagging a release. Package versions remain development identifiers, not a promise of semver stability.

No live provider credentials, AWS provisioning, external messages or model-controlled desktop actions are needed for the deterministic regression suites.

10 KiB Raw Permalink Blame History Unescape Escape