Waza v0.42.0: verifiable production integration
Released November 21, 2026.
Waza v0.42.0 adds the production-integration layer above the exact runner contract and durable execution shipped in v0.40.0 and v0.41.0. It adds signed execution provenance, component inventories, correlated telemetry, resource budgets, isolation capability declarations, portable adapters and an evidence-linked operator assurance package. These are bounded contracts, not a universal deployment or security approval.
The v0.42.0 tag carries 0.42.0 in version.txt and extension.yaml. This
article is a release record, not itself evidence of published binaries,
extension registry synchronization, a public image or attestations; verify those
independently. Use the
current published standalone and azd releases.
The download reference identifies the tag’s standalone assets separately from
extension registry synchronization and native execution evidence.
Source scope and release admission
Section titled “Source scope and release admission”This inventory covers the v0.42.0 milestone (milestone 2), whose issues were
closed when inspected on November 20, 2026. The merged-main source
frontier is 2dd5a32a20ffbb1ebc421e870386d9ab109e5fdc, including the readiness fixes in
PR #1084 and the opt-in keyless
execution signer in PR #1086.
Use the source comparison
and changelog
for the complete change history, including dependency and documentation updates.
Issue closure is not evidence of candidate qualification or publication.
Roadmap #756 ties the seven production-integration contracts together. The release procedure requires source-bound evidence; a different or stale green run is not sufficient.
Publication record
Section titled “Publication record”- The standalone v0.42.0 release
is tagged at
0dffb882834bedd3c798ca9d2917a2bf20b353cfand contains six binaries and checksums.txt. The separate azd extension release contains six platform packages andrelease-provenance.txt. - Release run 44981911687 published both releases but concluded with a failure: GitHub Actions policy denied its registry PR creation. A maintainer opened release-sync PR #1104 manually.
- The Release run skipped its downstream native package smoke. Native-platform run 44986511723 passed its native smoke matrix using a same-source rebuild, not the published bytes.
Production-integration contracts
Section titled “Production-integration contracts”| Milestone work | Included source behavior | Adoption and verification boundary |
|---|---|---|
| #769 execution provenance, PR #1000, PR #1086 | runner provenance generate, verify and keygen; in-toto v1 statements in DSSE bind canonical inputs, API/Waza/image identities, policy, runtime/model claims, artifact chain and terminal outcome. Opt-in --signer sigstore-keyless signs in GitHub Actions. Lab verifies local-key bundles. |
Local Ed25519 remains the default; its trust anchors must be supplied out of band. Keyless signing is optional and needs a verifier-supplied trusted root plus issuer/identity policy. Managed-KMS signing is unavailable (exit 5). GitHub build attestations are separate, not execution-outcome signatures. |
| #770 component inventory, PR #1008 | runner inventory generate and runner inventory verify, deterministic CycloneDX 1.6 component inventory bound to image/plugin digests, provenance bindings and Lab verification. |
Verify signed inventory and the actual consumed digest. Publication is restricted to the qualified tag-release path; previously published images are unqualified until republished and verified. Inventories are not vulnerability or license assessments. |
| #771 correlated execution telemetry, PR #997 | runner telemetry derives deterministic trace/span IDs, bounded metrics/logs and external scheduler correlation from retained runner documents; Lab exposes correlation. |
Signals are post-run, payload-free by default, with optional digest/length surrogates. OTLP/JSON files and OTLP/HTTP JSON are supported; live hot-path emission, OTLP gRPC/protobuf and raw payload export are not. |
| #772 resource budgets, PR #1002 | Optional runner request budgets, pre-execution capability admission, terminal budget reports, runner budgets and Lab inspection. |
Wall time/model turns/tool calls are enforced. CPU and supported peak-memory usage are measured, not enforced. Disk/process count and Windows peak memory are unsupported; reject is default, report explicitly allows unenforced limits. Older runners reject budget-bearing requests. |
| #773 portable isolation, PR #999 | Requested/available/effective capability declarations, isolation preflight, run admission and Lab diagnostics probe boundary violations. |
Waza enforces no sandbox itself. Native sandbox provider work #544/#545 is not in v0.42.0. Host probes cannot prove filesystem, secret or subprocess confinement; off-host network proof requires no external interfaces and does not cover loopback. |
| #774 portable adapters, PR #1001 | runner adapter translate and runner adapter dispatch map one vendor-neutral descriptor to one runner submit/checkpoint/lineage call; Lab translates offline. |
Request/policy bytes are preserved; ambiguous, lossy or scheduler input is rejected. External schedulers own queueing, placement, retries and fleets. Operators configure the bounded provenance-verifier hook and its trust policy; dispatch alone does not verify provenance. |
| #775 assurance package, PR #1023 | Versioned production-integration assurance 1.0 package, offline runner assurance build, generated conformance matrix, threat model, readiness checklist, migration and evidence links. |
Candidate guidance records unavailable states and links existing Lab views; it adds no new dashboard view and is not a deployment-specific approval. Reassess the exact released version and image digest. |
Read the provenance, inventory, telemetry, budgets, isolation, adapters and assurance package guides for contracts, safe examples, verification and troubleshooting.
Opt-in keyless execution signing
Section titled “Opt-in keyless execution signing”PR #1086 adds opt-in
--signer sigstore-keyless for runner provenance generate. It uses GitHub
Actions OIDC with public-good Fulcio and Rekor v1, and saves a Sigstore bundle
beside the unchanged DSSE statement. Keyless signing is optional and is not a
release requirement; local Ed25519 remains the default signer. Signing and
offline verification need a verifier-supplied Sigstore trusted root, and
verification also needs the exact OIDC issuer and workflow identity policy.
Not included: managed-KMS signing (still unavailable), certificate-transparency SCT and revocation checks, Lab/API keyless verification, GitHub Enterprise Server or custom OIDC issuers, and live public-good acceptance testing. See the execution provenance guide for permissions, egress, trust-root acquisition and public-log disclosure.
Evaluation and evidence fixes
Section titled “Evaluation and evidence fixes”| Milestone issue | Included fix |
|---|---|
| #1071 inline prompt input, commit 7cb755f | Independent inline prompt judges receive candidate output and task input; missing verdicts no longer claim all prompts passed. Task prompt graders bypass stale cached outcomes. |
| #1027 pairwise grading, PR #1081 | Successful skills-enabled and baseline outputs are compared as pairs; relative judgments drive impact and exit status. Wins/ties/losses and position inconsistencies are recorded. Offline saved-result pairwise regrading is rejected when baseline evidence is unavailable. |
| #1024 skill-disabled baseline, PR #1073 | Task overrides and ambient skills/agent definitions no longer leak into the disabled baseline. Task-only skill selection still runs the comparison. |
| #1025 consumed-input cache keys, PR #1074 | Consumed skill, instruction and fixture bytes affect cache identity. Unreadable consumed inputs run uncached; old entries miss once after upgrade. |
| #1029 changing task sets, PR #1075 | JSON comparisons preserve unavailable scores, rates and deltas as null with reasons, rather than inventing zero or emitting non-finite values. |
| #1030 strict replay, PR #1080 | Offline replay validates stored event order and supported status consistency; absent historical validation evidence remains unassessed. |
| #1057 repository capture, PR #1061 | Output-only runs can disable workspace capture only after admission proves no consumer needs file evidence. Run/trigger/optimizer and cache modes agree; capture stays enabled by default. |
| #1056 infrastructure errors, PR #1060 | Lab shows allowlisted per-trial infrastructure diagnostics even without grader results, without exposing raw session/responder/input errors. Closed release-artifact exports retain their existing contract. |
| #898 frozen judge evidence, PR #1065 | Opt-in judge-evidence 1.1 preserves declared fixture literals through capture/regrade. Undeclared sensitive values remain redacted; fixture declarations are not public evidence. |
| #964 Lab checkpoint validation, PR #990 | The browser matches native canonical report validation; unverifiable evidence remains explicitly unverified. |
| #984 orphan journal race, PR #1072 | The Linux regression waits for observed owner journal state before asserting orphan/retry behavior. |
Qualification, CI and performance
Section titled “Qualification, CI and performance”- #905 native platform evidence (PR #1007) separates native, cross-built, emulated and unavailable evidence for six targets. Native package smoke checks exact distributed binaries/extension packages and rejects architecture/digest mismatches and runtime launches. Published-package smoke is reported evidence, not yet a publication gate.
- #937 immutable release retry (PR #1076) reuses an existing registry-sync branch only when parent/tree match the validated sync; it never force-pushes a different tip.
- #981 completeness gate follow-ups (PR #1003) cover PR retargeting, source-bound receipt races and consolidated Docker validation.
- #974 dashboard performance (PR #1005) restores the retry-disabled Playwright wall-time budget with lazy Ajv compilation and direct-route rendering.
- #835 container Go caches (PR #836) reuses module/build caches without changing the runtime contract. Related measured CI work under #825 balances native and remaining-package shards; this does not claim that the broader open issue is complete.
- #756 qualification adds exact-current-main candidate qualification with binaries, OCI image archive, CycloneDX inventory, GitHub provenance/SBOM attestations and offline verification. Candidate runs publish no tag, release or registry image. The tag path publishes the qualified artifacts without rebuilding binaries. Readiness fixes add public-adapter signed-provenance verification and decouple native probe behavior from a hard-coded bundled CLI version (#876).
Other included source changes
Section titled “Other included source changes”Product usage telemetry #1045
(PR #1069) is separate from
runner execution telemetry: allowlisted pseudonymous events are collected by
default on developer machines, never in CI or development builds, with
30-day retention. Every opt-out wins: --no-telemetry,
WAZA_TELEMETRY_OPTOUT=1, DO_NOT_TRACK=1, waza telemetry disable,
project denial and azd denial. See the usage telemetry guide.
Dependency updates include Copilot Go SDK 1.0.34 and bundled CLI 1.0.112, coordinated site/dashboard updates, and expanded native shard coverage. Documentation/sample readiness fixes align production commands, examples, provenance hook responsibilities and the generated assurance page with this candidate. Historical v0.41.0 publication records and pinned downloads remain unchanged.
Release media
Section titled “Release media”No v0.42.0 screenshots or videos are part of this release record. Any media is tracked separately by the planning-only #1085 v0.42.0 media task; no capture or publication is authorized by this article. Follow the release-media procedure after exact-source scope and deterministic capture paths are verified; drafts are not publication permission.