Production integration assurance package
Package 1.0 · v0.42.0 · candidate. Evidence-linked qualification guidance for the v0.42.0 production-integration contracts. This package describes the checked-in candidate; the qualification workflow separately binds its exact run and outputs. Neither is a blanket readiness certificate or deployment-specific risk assessment.
This is an operator qualification aid, not a blanket production-readiness guarantee or a substitute for deployment-specific risk assessment. Evidence for the #769–#774 contracts is linked below; this package adds no new Lab dashboard view.
Source evidence links are pinned to the v0.42.0 release tag. The source package remains a candidate until tag-release qualification completes; its offline builder verifies each evidence file against the checked-out source.
Contents: Conformance matrix · Threat model · Readiness checklist · Verification limits · Migration and reassessment · Existing Lab evidence
Operator conformance matrix
Section titled “Operator conformance matrix”| Capability | State | Claim and verification | Declared limitations | Evidence |
|---|---|---|---|---|
Execution provenance — local signing and offline verification (#769)provenance-local-signing |
supported |
The provenance 1.0 profile binds canonical request identity, runner and Waza claims, policy, artifact manifest and objects, terminal outcome, and digest-bound extensions. This build generates and verifies local Ed25519 signatures offline against verifier-supplied public keys. Use waza runner provenance generate and waza runner provenance verify; the verifier recomputes claims from bundle members and requires the trust anchor out of band. |
This is Waza execution provenance; a signature does not establish who controls a locally generated key unless operators distribute and protect trust anchors. The release workflow separately uses GitHub artifact attestations for build outputs, which do not replace the Waza-specific claims. | Provenance contract and signing modes Provenance implementation tests CLI verification tests Contracts: provenance-bundle-1.0.schema.json, provenance-statement-1.0.schema.json, provenance-verification-1.0.schema.json |
Execution provenance — keyless and managed KMS (#769, #756)provenance-keyless-kms |
unavailable |
The Waza execution-provenance CLI has no Sigstore keyless or managed-KMS signer. Requesting either Waza signer returns the documented unavailable state. Separately, the release qualification workflow produces GitHub artifact attestations for exact binaries and the OCI image digest and verifies their bundles offline. | Waza execution provenance supports only local Ed25519 signing; it does not provide external identity, KMS, keyless signing, key rotation, or managed trust distribution. GitHub artifact attestations cover build outputs and do not sign Waza execution outcomes. | Unavailable signing modes Provenance CLI tests Contracts: provenance-verification-1.0.schema.json |
Runner image and plugin component inventory (#770)image-inventory-generation-verification |
supported |
The inventory 1.0 profile generates deterministic CycloneDX component inventories and verifies their subject, components, and immutable image or plugin digest offline. Qualification attests the inventory to the exact OCI manifest digest; the tag-release publisher preserves that digest and publishes the inventory attestation. Generate an inventory from the qualified image’s binary and OS package inputs, run waza runner inventory verify against its OCI manifest digest, and verify both the workflow-run bundle and registry-attached release attestation offline. |
Inventory is evidence, not vulnerability or license analysis. A digest check alone does not authenticate an inventory; verify its attestation or signed provenance binding. | Inventory format and verification Inventory implementation tests Reference image publisher Contracts: component-inventory-1.0.schema.json, inventory-verification-1.0.schema.json |
Runner image inventory — refresh and previously published images (#770, #756)inventory-refresh-and-legacy-images |
unavailable |
This build does not retroactively attest images published before the inventory workflow was introduced; registry publication occurs only for a qualified versioned release rather than as an automatic refresh guarantee. Inspect the tag-release reference-image publication workflow and require an inventory attestation for the exact image digest being admitted. | Previously published image digests remain without inventory attestations; this workflow does not replace old version tags. Candidate qualification retains inventory and attestation evidence in workflow-run artifacts only. Do not infer registry publication or inventory coverage from a mutable tag or an unrelated workflow run. | Release-only reference image publisher and attestation steps Inventory verification requirements Contracts: component-inventory-1.0.schema.json |
Execution telemetry — post-run derivation and export (#771)post-run-telemetry |
supported |
The telemetry 1.0 report derives correlated traces, metrics, and logs from completed runner contract documents. It exports OTLP/JSON files or sends OTLP/HTTP JSON; payloads are excluded by default. Run waza runner telemetry with a terminal and optional matching manifest, attempt, lineage, result, or aggregate, then inspect the versioned report and export diagnostics. |
This is post-run derivation, not live emission from the execution path. It requires completed contract documents; OTLP gRPC/protobuf and raw-payload export are not provided. | Telemetry inputs, payload policy, and export modes Telemetry derivation tests Telemetry CLI tests Contracts: execution-telemetry-1.0.schema.json |
Execution budgets — wall time, model turns, and tool calls (#772)wall-time-and-activity-budgets |
supported |
Wall-time deadlines and model-turn/tool-call starts are checked and enforced before or during a selected attempt on supported hosts; unsupported limits can be rejected before execution. Inspect waza runner budgets, use unsupported: reject, and verify the terminal’s configured limits, usage, enforcement, and termination report. |
Counted limits can cancel after the over-limit call has begun. Budgets are controls, not a sandbox, and are not per-model/provider billing guarantees. | Budget host capability matrix Budget enforcement and report tests Runner budget CLI tests Contracts: execution-budget-1.0.schema.json, runner-request-1.0.schema.json, runner-terminal-1.0.schema.json |
Execution budgets — CPU, memory, disk, and process limits (#772, #756)os-resource-budget-gaps |
unsupported |
This runner does not enforce CPU, memory, disk, or process-count quotas. CPU and peak RSS are measured only at runner-process scope on supported OSes; disk and process count are not measured. Inspect each OS row in waza runner budgets; any unsupported configured limit must be rejected or explicitly reported as unenforced. |
Process-scoped CPU/RSS observations are not isolated per attempt and can include concurrent work or pre-attempt usage. No cgroup, job-object, rlimit, workspace-disk quota, or process-tree accounting is claimed. | OS-specific budget measurements and unsupported limits Budget capability implementation Contracts: execution-budget-1.0.schema.json |
Isolation — off-host network and remote-tool boundaries (#773)off-host-network-and-remote-tools |
supported |
The host provider can prove off-host network and remote-tool unreachability only when the process has no up non-loopback interface and the required probes support that result. Loopback remains within the boundary. Run isolation preflight in the actual worker boundary with representative declared remote endpoints; accept only effective: restricted backed by the required host probes. |
Waza declares and probes but does not enforce isolation. The proof covers off-host reachability only under the documented no-interface condition; it does not restrict loopback or prove general sandboxing. | Network and remote-tool probe limits Isolation provider implementation Isolation probe tests Contracts: execution-isolation-1.0.schema.json |
Isolation — filesystem, environment, subprocess, and native sandbox (#773, #756)filesystem-environment-subprocess-isolation |
unavailable |
This build cannot prove restricted filesystem writes, secret withholding, process-wide subprocess denial, or a Copilot-native sandbox. A host declaration is a hypothesis, not proof. Treat requested dimensions as unavailable unless an external provider supplies independently verifiable enforcement evidence; check the preflight diagnostics before dispatch. | Waza itself enforces no isolation dimension. Filesystem denials, absent secret names, and sampled subprocess denials are inconclusive; the native sandbox provider is absent. | Isolation dimensions and unavailable providers Isolation provider implementation Isolation conformance tests Contracts: execution-isolation-1.0.schema.json |
External execution adapter — translation and contract reply mapping (#774)adapter-translation |
supported |
An adapter descriptor is translated to exactly one public runner contract call or rejected, and the reply maps to a status and returned manifest without re-encoding request or policy bytes. Run the vendor-neutral CLI and container fixtures for submit, reuse, retry, checkpoint, resume, lineage, malformed descriptors, and returned manifest mapping. | The adapter makes one call per descriptor and has no built-in queue, placement, lease, retry loop, fleet manager, or scheduler. Those guarantees belong to the external system. | Adapter public contract and translation-only boundary Adapter implementation and fixture tests CLI dispatch sequence tests Contracts: adapter-work-descriptor-1.0.schema.json, adapter-translation-1.0.schema.json, adapter-result-1.0.schema.json |
External execution adapter — scheduler orchestration (#774, #756)adapter-scheduler-orchestration |
unsupported |
No queue, placement, lease, retry loop, fleet manager, or production scheduler is included. Inspect the one-call adapter boundary and confirm that dispatch and redelivery policy remain with the external scheduler. | The adapter makes one public contract call per descriptor and does not claim delivery, at-least-once, exactly-once, or scheduling guarantees. | Translation-only scheduler boundary Adapter contract implementation Translation-only architecture test Contracts: adapter-result-1.0.schema.json |
External execution adapter — configured provenance hook (#774, #756)adapter-provenance-hook |
supported |
The adapter CLI can invoke a configured verifier after a terminal public response and records only its validated state; it never retries or changes the runner outcome based on that state. Run the adapter CLI with a test verifier executable; confirm it receives the returned manifest digest and public result, produces verified, and dispatch remains exactly once. |
The configured verifier is an operator-trusted executable. The adapter does not implement a scheduler or independently authenticate the verifier’s assertion. | Verifier command contract and CLI flags Adapter verifier implementation Configured CLI verifier E2E test Contracts: adapter-result-1.0.schema.json |
Threat model
Section titled “Threat model”| Boundary | Threat | Current controls | Residual risk | Evidence |
|---|---|---|---|---|
| Runner request, artifact store, terminal, and provenance bundle | An attacker or storage fault alters inputs, artifacts, outcomes, or adds unbound bundle members after execution. | Provenance verification recomputes canonical identities and outcome bindings, checks digests and signatures, and rejects unbound or irregular bundle members. Inventory and terminal readers validate their closed profiles. | Verification is only as trustworthy as the verifier binary, the operator-supplied key, and the storage from which the original expected digest was obtained. | Offline provenance verification Provenance verifier tests Inventory verification |
| Signing identity and verifier trust store | A compromised or misidentified local key creates signatures that an operator mistakes for trusted publisher provenance. | The verifier accepts only public keys supplied out of band and checks the signed statement against bundle members. | Waza local-key generation, custody, distribution, rotation, identity binding, keyless signing, and managed KMS remain operator responsibilities or unavailable. GitHub artifact attestations establish workflow identity for build evidence, not execution outcomes. | Signing modes and trust-anchor requirements Keyless and KMS unavailable state |
| Reference image registry and consumed image digest | A mutable tag or previously published image is treated as covered by an inventory attestation that was never produced for its digest. | Inventory verification pins the consumed immutable digest; the tag-release workflow emits an inventory attestation after exact-source qualification. | Prior image digests have no inventory attestation; candidate qualification does not publish an image. Inventory does not evaluate vulnerabilities or licenses. | Reference image workflow Digest-bound inventory requirements |
| Waza process, runtime, filesystem, environment, subprocesses, and tools | Untrusted work escapes its workspace, reads credentials, starts unauthorized child processes, or reaches a remote service. | Isolation preflight reports only evidenced capabilities and rejects requested restricted dimensions it cannot prove. Off-host network and remote-tool checks require no up non-loopback interface. | Waza does not enforce isolation. Filesystem, environment, and subprocess restriction remain unproven; loopback is reachable; the native sandbox provider is absent. | Isolation threat boundaries and probe semantics Isolation providers |
| Runner process and host resource scheduler | A workload exceeds expected compute, memory, storage, time, or child-process usage. | Wall-time, model-turn, and tool-call limits are enforced where documented. Unsupported requested limits can be rejected before execution and the terminal reports measured or unavailable states. | CPU, memory, disk, and process-count quotas are not enforced. Process-scope CPU/RSS measurements are not per-attempt; disk and process count are unmeasured. | Budget enforcement and OS accounting matrix Budget capability implementation |
| Completed execution evidence and telemetry receiver | Identifiers or operational metadata are disclosed to a telemetry backend, or payloads are unintentionally exported. | Telemetry is derived after execution, excludes payloads by default, catalogs attributes, and bounds signals and dimensions. | Exporting identifiers to a receiver is an operator choice. There is no live execution-path emission, and digest policy still exposes payload-derived hashes and lengths. | Telemetry payload policy and export Telemetry export tests |
| External scheduler, adapter invocation, and Waza public runner contract | An operator mistakes one translation call for durable queueing, safe automatic retries, exclusive scheduling, or provenance verification. | Adapter rejects lossy scheduler fields, makes one contract call per descriptor, and maps the actual reply; separate public submission policies cover explicit reuse or distinct attempts. | The external scheduler owns delivery, retry, placement, leases, duplicate handling, and trust policy. The optional provenance hook is not configured by the CLI. | Adapter operation and non-goals Translation-only architecture guard |
Production-readiness checklist
Section titled “Production-readiness checklist”- Record the exact Waza version, runner API profile, worker OS, runner image digest, and operator trust anchors for each deployment. — The evidence names immutable digests and the installed CLI’s observed version; do not treat a tag, request, or configured identity as proof of execution.
pin-and-identify. Evidence: Runner capabilities and identity contracts
Provenance verification and pinning - Verify the component inventory and its attestation against the exact image digest before admitting a reference image. — Keep the verifier report and attestation result for the digest actually consumed. For images without an inventory, republish explicitly or record that the evidence is unavailable.
verify-image-inventory. Evidence: Inventory generation, verification, and image attestation
Manual reference-image publication workflow
Operator example - Generate and verify execution provenance for completed work, with a locally managed Ed25519 key and a public trust anchor distributed out of band. — Require a verified report bound to the exact request, image, artifact manifest, artifact objects, and terminal outcome. Record unavailable keyless/KMS signing as a deployment gap.
verify-provenance. Evidence: Provenance CLI and verification procedure
Provenance contract schemas - Inspect the budget matrix on every worker OS and set unsupported policy to reject when an unenforceable limit would make the work unsafe. — Accept only limits shown enforced by that host. Treat CPU/RSS as runner-process measurements where available, not per-attempt quotas; do not claim disk or process-count accounting.
check-host-budgets. Evidence: Per-OS budget capability matrix
Budget example request - Enforce the execution boundary outside Waza and run isolation preflight inside the final worker/container boundary before dispatch. — Require evidenced off-host network/remote-tool restrictions under the no-non-loopback-interface condition. Do not infer filesystem, environment, or subprocess isolation from host declarations or inconclusive probes.
prove-external-isolation. Evidence: Isolation preflight and provider limits
Isolation contract schema - Derive telemetry only from completed contract evidence and choose an approved OTLP/HTTP receiver and payload policy. — Confirm the receiver, access controls, retention, and identifier handling; use the default payload exclusion unless digest disclosure is explicitly accepted.
export-telemetry-deliberately. Evidence: Post-run telemetry inputs and export behavior
Telemetry 1.0 schema - Keep queueing, placement, leases, retry policy, and delivery guarantees in the external scheduler; dispatch through one adapter call per descriptor. — Exercise the external scheduler’s failure, duplicate-delivery, and retry behavior independently. Treat adapter provenance
not_configuredas no verification.integrate-adapter. Evidence: Adapter contract and limitations
Portable adapter example - Use the existing Lab panels to inspect published evidence and diagnostics; do not assume the assurance package adds a new dashboard view. — Correlate each view with its source contract and preserve the JSON/report used for the operator decision.
inspect-existing-lab-evidence. Evidence: Existing Lab evidence and navigation
Versioned assurance package - After every Waza or runner-image upgrade, re-negotiate capabilities, re-verify the consumed image inventory, and reassess the deployment against this package version. — Compare the new schemas and package applicability to the previous evidence; regenerate signed evidence for the upgraded image and retain any unavailable or unsupported states.
reassess-after-upgrade. Evidence: Runner compatibility and migration
Versioned production package
Verification limits
Section titled “Verification limits”- This package describes the v0.42.0 source candidate and the listed public contract profiles; it does not certify every Waza binary, image, host, runtime, scheduler, model, or deployment.
- Waza local Ed25519 signatures are offline-verifiable only relative to a verifier-supplied trust anchor; Waza keyless Sigstore and managed-KMS signers remain unavailable. GitHub artifact attestations separately cover qualification outputs and are not Waza execution provenance.
- Candidate image and inventory evidence remains in workflow-run artifacts. Registry publication and inventory attestation occur only in the qualified tag-release path; previously published digests are not retroactively covered.
- Telemetry is derived from completed runner documents and can be exported as files or OTLP/HTTP JSON; there is no live execution-path telemetry emission.
- Budgets enforce wall time, model-turn, and tool-call limits only. CPU and memory observations are process-scoped on supported OSes and are not per-attempt limits; disk and process-count accounting are unavailable.
- Isolation evidence proves only off-host network and remote-tool boundaries when no up non-loopback interface is present. Loopback remains reachable; Waza does not enforce filesystem, environment, subprocess, or general sandbox boundaries.
- Adapters translate one descriptor to one public runner call. No scheduler is included; the optional provenance hook runs only when an operator configures a verifier.
- A green local test or documentation-site build establishes repository conformance only; it is not a deployment-specific assessment, penetration test, vulnerability scan, or production workload measurement.
Migration and reassessment
Section titled “Migration and reassessment”- Migration: The package and its evidence links are additive; they do not change runner request, terminal, manifest, or adapter wire contracts.
- Older runners: Older runners do not implement these v0.42.0 features and may reject newly added contract fields. Negotiate the installed runner’s capabilities and upgrade deliberately before depending on an optional profile.
- Previously published images: Images published before inventory attestation remain uncovered; the tag-release publisher refuses to replace an existing versioned image tag, and existing image digests are not retroactively covered.
- After upgrades: For each Waza or image upgrade, verify the exact consumed digest, refresh inventory and provenance evidence, repeat host capability/preflight checks, and review updated limitations rather than inheriting a prior qualification.
Existing Lab evidence
Section titled “Existing Lab evidence”The following existing evidence views are linked rather than duplicated. Use the Dashboard Lab guide for navigation and local Lab setup.
| Issue | Existing Lab panel | Operator guide |
|---|---|---|
| #769 | Execution provenance | Execution provenance |
| #770 | Component inventory | Component inventory |
| #771 | Execution telemetry | Execution telemetry |
| #772 | Execution budgets | Execution budgets |
| #773 | Execution isolation | Execution isolation |
| #774 | External adapter dispatch | External adapter dispatch |
Maintainers: update the versioned source package at docs/PRODUCTION_INTEGRATION_ASSURANCE-1.0.json, then run go run ./cmd/waza runner assurance build --format markdown from the repository root and commit the regenerated page with the source change.