openclaw-testing

name	openclaw-testing
description	Choose, run, rerun, or debug OpenClaw tests, CI checks, Docker E2E lanes, release validation, and the cheapest safe verification path.

OpenClaw Testing

Use this skill when deciding what to test, debugging failures, rerunning CI, or validating a change without wasting hours.

Read First

• docs/reference/test.md for local test commands.
• docs/ci.md for CI scope, release checks, Docker chunks, and runner behavior.
• Scoped AGENTS.md files before editing code under a subtree.

Default Rule

Prove the touched surface first. Do not reflexively run the whole suite.

1. Inspect the diff and classify the touched surface:
   • normal source checkout, source change: pnpm changed:lanes --json, then pnpm check:changed
   • normal source checkout, tests only: pnpm test:changed
   • normal source checkout, one failing file: pnpm test <path-or-filter> -- --reporter=verbose
   • Codex worktree or linked/sparse checkout, one/few explicit files: node scripts/run-vitest.mjs <path-or-filter>
   • Codex worktree or linked/sparse checkout, changed gates or anything broad: node scripts/crabbox-wrapper.mjs run --provider blacksmith-testbox ... --shell -- "pnpm check:changed"
   • workflow-only: git diff --check, workflow syntax/lint (actionlint when available)
   • docs-only: pnpm docs:list, docs formatter/lint only if docs tooling changed or requested
2. Reproduce narrowly before fixing.
3. Fix root cause.
4. Rerun the same narrow proof.
5. Broaden only when the touched contract demands it.

Guardrails

• Do not kill unrelated processes or tests. If something is running elsewhere, treat it as owned by the user or another agent.
• Do not run expensive local Docker, full release checks, full pnpm test, or full pnpm check unless the user asks or the change genuinely requires it.
• Prefer GitHub Actions for release/Docker proof when the workflow already has the prepared image and secrets.
• Use scripts/committer "<msg>" <paths...> when committing; stage only your files.
• If deps are missing, run pnpm install, retry once, then report the first actionable error.
• In a Codex worktree or linked/sparse checkout, do not run direct local pnpm test*, pnpm check*, pnpm crabbox:run, or scripts/committer until you have verified pnpm will not reconcile or reinstall dependencies. Use node scripts/run-vitest.mjs for tiny local proof, node scripts/crabbox-wrapper.mjs for Testbox, and git commit --no-verify only after the relevant remote or node-wrapper proof is already clean.
• For Blacksmith Testbox proof, use Crabbox first. pnpm crabbox:run -- --provider blacksmith-testbox --timing-json -- <command...> warms, claims, syncs, runs, reports, and cleans up one-shot boxes. Reuse only an id/slug created in this operator session; blacksmith testbox list is diagnostics only, not a shared work queue.

Local Test Shortcuts

pnpm changed:lanes --json
pnpm check:changed       # changed typecheck/lint/guards; no Vitest
pnpm test:changed        # cheap smart changed Vitest targets
OPENCLAW_TEST_CHANGED_BROAD=1 pnpm test:changed
pnpm test <path-or-filter> -- --reporter=verbose
OPENCLAW_VITEST_MAX_WORKERS=1 pnpm test <path-or-filter>

Use targeted file paths whenever possible. Avoid raw vitest; use the repo pnpm test wrapper so project routing, workers, and setup stay correct. When the checkout is a Codex worktree, prefer the direct node harness instead:

node scripts/run-vitest.mjs <path-or-filter>

That keeps the test scoped without giving pnpm a chance to run dependency status checks or install reconciliation in a linked worktree.

Command Semantics

• pnpm check and pnpm check:changed do not run Vitest tests. They are for typecheck, lint, and guard proof.
• pnpm test and pnpm test:changed run Vitest tests.
• pnpm test:changed is intentionally cheap by default: direct test edits, sibling tests, explicit source mappings, and import-graph dependents.
• OPENCLAW_TEST_CHANGED_BROAD=1 pnpm test:changed is the explicit broad fallback for harness/config/package edits that genuinely need it.
• Do not run extension sweeps just because core changed. If a core edit is for a specific plugin bug, run that plugin's tests explicitly. If a public SDK or contract change needs consumer proof, choose the smallest representative plugin/contract tests first, then broaden only when the risk justifies it.
• The test wrapper prints a short [test] passed|failed|skipped ... in ... line. Vitest's own duration is still the per-shard detail.

Routing Model

• pnpm changed:lanes --json answers "which check lanes does this diff touch?" It is used by pnpm check:changed for typecheck/lint/guard selection.
• pnpm test:changed answers "which Vitest targets are worth running now?" It uses the same changed path list, but applies a cheaper test-target resolver.
• Direct test edits run themselves. Source edits prefer explicit mappings, sibling *.test.ts, then import-graph dependents. Shared harness/config/root edits are skipped by default unless they have precise mapped tests.
• Shared group-room delivery config and source-reply prompt edits are precise mapped tests: they run the core auto-reply regressions plus Discord and Slack delivery tests so cross-channel default changes fail before a PR push.
• Public SDK or contract edits do not automatically run every plugin test. check:changed proves extension type contracts; the agent chooses the smallest plugin/contract Vitest proof that matches the actual risk.
• Use OPENCLAW_TEST_CHANGED_BROAD=1 pnpm test:changed only when a harness, config, package, or unknown-root edit really needs the broad Vitest fallback.

CI Debugging

Start with current run state, not logs for everything:

gh run list --branch main --limit 10
gh run view <run-id> --json status,conclusion,headSha,url,jobs
gh run view <run-id> --job <job-id> --log

• Check exact SHA. Ignore newer unrelated main unless asked.
• For cancelled same-branch runs, confirm whether a newer run superseded it.
• Fetch full logs only for failed or relevant jobs.

GitHub Release Workflows

Use the smallest workflow that proves the current risk. The full umbrella is available, but it is usually the last step after narrower proof, not the first rerun after a focused patch.

Full Release Validation

Full Release Validation (.github/workflows/full-release-validation.yml) is the manual "everything before release" umbrella. It resolves a target ref, then dispatches:

• manual CI for the full normal CI graph, with Android enabled via include_android=true
• Plugin Prerelease for release-only plugin static checks, extension shards, the release-only agentic-plugins shard, and plugin product Docker lanes
• OpenClaw Release Checks for install smoke, cross-OS release checks, live and E2E checks, Docker release-path suites, OpenWebUI, QA Lab, fast Matrix, and Telegram release lanes
• optional post-publish Telegram E2E when a package spec is supplied

Run it only when validating an actual release candidate, after broad shared CI or release orchestration changes, or when explicitly asked:

gh workflow run full-release-validation.yml \
  --repo openclaw/openclaw \
  --ref main \
  -f ref=<branch-or-sha> \
  -f provider=openai \
  -f mode=both \
  -f release_profile=stable

Run the workflow itself from the trusted current ref, normally --ref main; child workflows are dispatched from that same ref even when ref points at an older release branch or tag. Full Release Validation has no separate child workflow ref input; choose the trusted harness by choosing the workflow run ref. Use release_profile=minimum|stable|full to control live/provider breadth: minimum keeps the fastest OpenAI/core release-critical set, stable adds the stable provider/backend set, and full adds the broad advisory provider/media matrix. Do not make full faster by silently dropping suites; optimize setup, artifact reuse, and sharding instead. The parent verifier job appends a child overview plus slowest-job tables for child runs; rerun only that verifier after a child rerun turns green.

Standalone manual CI dispatches do not run the plugin prerelease suite, the extension batch sweep, or the release-only agentic-plugins Vitest shard. Those lanes are intentionally reserved for the separate Plugin Prerelease child so PRs, main pushes, and ad hoc broad CI checks do not spend Docker/package time or all-plugin runtime time on release-only product coverage.

If a full run is already active on a newer origin/main, prefer watching that run over dispatching a duplicate. Do not cancel release, release-check, or child workflow runs unless Peter explicitly asks for cancellation.

The child-dispatch jobs record the child run ids. The final Verify full validation job re-queries those child runs and is the canonical parent gate. If a child workflow failed but was later rerun successfully, rerun only the failed parent verifier job; do not dispatch a new full umbrella unless the release evidence is stale.

For bounded recovery after a focused fix, pass -f rerun_group=<group>. Supported umbrella groups are all, ci, plugin-prerelease, release-checks, install-smoke, cross-os, live-e2e, package, qa, qa-parity, qa-live, and npm-telegram. Use the narrowest group that covers the failed box. After a targeted release-check fix, do not restart the full umbrella by habit: dispatch the matching rerun_group and rerun only the parent verifier/evidence step after the child is green unless the release evidence is stale. For a single failed live/E2E shard, use -f rerun_group=live-e2e -f live_suite_filter=<suite_id> so the Blacksmith workflow only spends setup and queue time on that suite.

Release Evidence

After release-candidate validation or before a release decision, record the important run ids in the private openclaw/releases-private evidence ledger. Use the manual OpenClaw Release Evidence (openclaw-release-evidence.yml) workflow there. It writes durable summaries under evidence/<release-id>/ and commits:

• release-evidence.md
• release-evidence.json
• index.json
• runs/<label>.json

Use one run per line:

full-release-validation openclaw/openclaw <run-id> blocking
package-acceptance openclaw/openclaw <run-id> blocking
release-checks openclaw/openclaw <run-id> blocking

Store summaries, run URLs, artifact metadata, timings, pass/fail state, and short release-manager notes there. Do not store raw logs, provider prompts/responses, channel transcripts, signing material, or secret-bearing config in git; raw logs stay in Actions artifacts.

When Full Release Validation completes and OPENCLAW_RELEASES_PRIVATE_DISPATCH_TOKEN is configured in the public repo, it requests the private OpenClaw Release Evidence From Full Validation workflow. That private workflow reads the parent full-validation run, extracts the child CI/release-checks/Telegram run ids from the parent logs, and opens the evidence PR automatically. If the token is absent or the run predates this wiring, trigger that private workflow manually with the full-validation run id.

Release Checks

OpenClaw Release Checks (openclaw-release-checks.yml) is the release child workflow. It is broader than normal CI but narrower than the umbrella because it does not dispatch the separate full normal CI child. It runs Package Acceptance with artifact-native delta lanes and telegram_mode=mock-openai, so the release package tarball also goes through offline plugin proof, bundled-channel compat, and Telegram package QA. The Docker release-path chunks cover the overlapping package/update/plugin lanes. Use it when release-path validation is needed without rerunning the entire umbrella.

gh workflow run openclaw-release-checks.yml \
  --repo openclaw/openclaw \
  --ref main \
  -f ref=<branch-or-sha> \
  -f provider=openai \
  -f mode=both \
  -f release_profile=stable \
  -f rerun_group=all

Release-check rerun groups are all, install-smoke, cross-os, live-e2e, package, qa, qa-parity, and qa-live. OpenClaw Release Checks uses the trusted workflow ref to resolve the selected ref once as release-package-under-test and passes that artifact into cross-OS release checks, release-path Docker live/E2E checks, and Package Acceptance. When Full Release Validation dispatches release checks, it passes the requested branch/tag plus an expected_sha so branch/tag refs resolve through the fast remote-ref path while the package and QA jobs still validate the exact SHA.

The full install-smoke child is split on purpose: one job prepares or reuses the target-SHA GHCR root Dockerfile smoke image, QR package install runs in its own job, root Dockerfile/gateway smokes pull the prepared image, and installer/Bun smokes pull the same image while building only their small installer images. If install-smoke gets slow again, first check whether the root image was reused or rebuilt before adding/removing coverage.

The full-profile native live media shards use the prebuilt ghcr.io/openclaw/openclaw-live-media-runner:ubuntu-24.04 container so ffmpeg/ffprobe are already present. If those jobs suddenly spend minutes in dependency setup again, first check the Live Media Runner Image workflow and the Verify preinstalled live media dependencies step before assuming the media tests themselves slowed down.

The release Docker path intentionally shards the plugin/runtime tail. The workflow uses plugins-runtime-plugins, plugins-runtime-services, and plugins-runtime-install-a through plugins-runtime-install-d; aggregate aliases such as plugins-runtime-core, plugins-runtime, and plugins-integrations remain for manual reruns.

The release QA parity box is internally split into candidate and baseline lane jobs, followed by a report job that downloads both artifacts and runs pnpm openclaw qa parity-report. For parity failures, inspect the failed lane first; inspect the report job when both lane summaries exist but the comparison fails.

QA Lab Matrix Profiles

pnpm openclaw qa matrix defaults to --profile all. Do not assume the CLI default is the fast release path. Use explicit profiles:

• --profile fast: release-critical Matrix transport contract; add --fail-fast only when the target CLI supports it
• --profile transport|media|e2ee-smoke|e2ee-deep|e2ee-cli: sharded full Matrix proof
• OPENCLAW_QA_MATRIX_NO_REPLY_WINDOW_MS=3000: CI-friendly no-reply quiet window when paired with fast or sharded gates

QA-Lab - All Lanes uses explicit fast Matrix on scheduled runs; manual dispatch keeps matrix_profile=all as the default and always shards that full Matrix selection. OpenClaw Release Checks uses explicit fast Matrix; run the all-lanes workflow when release investigation needs full Matrix media/E2EE inventory.

Reusable Live/E2E Checks

OpenClaw Live And E2E Checks (Reusable) (openclaw-live-and-e2e-checks-reusable.yml) is the preferred entry point for targeted live, Docker, model, and E2E proof. Inputs let you turn off unrelated lanes:

gh workflow run openclaw-live-and-e2e-checks-reusable.yml \
  --repo openclaw/openclaw \
  --ref main \
  -f ref=<sha> \
  -f include_repo_e2e=false \
  -f include_release_path_suites=false \
  -f include_openwebui=false \
  -f include_live_suites=true \
  -f live_models_only=true \
  -f live_model_providers=fireworks

Useful knobs:

• docker_lanes='<lane[,lane]>': run selected Docker scheduler lanes against prepared artifacts instead of the release chunk matrix. Multiple selected lanes fan out as parallel targeted Docker jobs after one shared package/image preparation step.
• include_live_suites=false: skip live/provider suites when testing Docker scheduler or release packaging only.
• live_models_only=true: run only Docker live model coverage.
• live_model_providers=fireworks (or comma/space separated providers): run one targeted Docker live model job instead of the full provider matrix.
• blank live_model_providers: run the full live-model provider matrix.

Release-path Docker chunks are currently core, package-update-openai, package-update-anthropic, package-update-core, plugins-runtime-plugins, plugins-runtime-services, plugins-runtime-install-a, plugins-runtime-install-b, plugins-runtime-install-c, plugins-runtime-install-d, bundled-channels-core, bundled-channels-update-a, bundled-channels-update-b, and bundled-channels-contracts. The aggregate bundled-channels, plugins-runtime-core, plugins-runtime, and plugins-integrations chunks remain valid for manual one-shot reruns, but release checks use the split chunks.

When live suites are enabled, the workflow shards broad native pnpm test:live coverage through scripts/test-live-shard.mjs instead of one serial live-all job:

• native-live-src-agents
• native-live-src-gateway-core
• native-live-src-gateway-profiles (release CI runs this with provider filters such as OPENCLAW_LIVE_GATEWAY_PROVIDERS=anthropic)
• native-live-src-gateway-backends
• native-live-test
• native-live-extensions-a-k
• native-live-extensions-l-n
• native-live-extensions-openai
• native-live-extensions-o-z
• native-live-extensions-o-z-other
• native-live-extensions-xai
• native-live-extensions-media
• native-live-extensions-media-audio
• native-live-extensions-media-music
• native-live-extensions-media-music-google
• native-live-extensions-media-music-minimax
• native-live-extensions-media-video

Use node scripts/test-live-shard.mjs <shard> --list to see the exact files before rerunning a failed native live shard. The aggregate o-z and media shards remain useful locally; release CI uses the smaller provider/media shards so one live-provider flake does not force a broad native live rerun.

For model-list or provider-selection fixes, use live_models_only=true plus the specific live_model_providers allowlist. Confirm logs show the expected OPENCLAW_LIVE_PROVIDERS and selected model ids before declaring proof.

Docker

Docker is expensive. First inspect the scheduler without running Docker:

OPENCLAW_DOCKER_ALL_DRY_RUN=1 pnpm test:docker:all
OPENCLAW_DOCKER_ALL_DRY_RUN=1 OPENCLAW_DOCKER_ALL_LANES=install-e2e pnpm test:docker:all
OPENCLAW_DOCKER_ALL_LANES=install-e2e node scripts/test-docker-all.mjs --plan-json

Run one failed lane locally only when explicitly asked or when GitHub is not usable:

OPENCLAW_DOCKER_ALL_LANES=<lane> \
OPENCLAW_DOCKER_ALL_BUILD=0 \
OPENCLAW_DOCKER_ALL_PREFLIGHT=0 \
OPENCLAW_SKIP_DOCKER_BUILD=1 \
OPENCLAW_DOCKER_E2E_BARE_IMAGE='<prepared-bare-image>' \
OPENCLAW_DOCKER_E2E_FUNCTIONAL_IMAGE='<prepared-functional-image>' \
pnpm test:docker:all

For release validation, prefer the reusable GitHub workflow input:

docker_lanes: install-e2e

Multiple lanes are allowed:

docker_lanes: install-e2e bundled-channel-update-acpx

That skips the release chunk matrix and runs one targeted Docker job against the prepared GHCR images and the selected package artifact. Rerun commands generated inside GitHub artifacts include package_artifact_run_id, package_artifact_name, docker_e2e_bare_image, and docker_e2e_functional_image when available, so failed lanes can reuse the exact tarball and prepared images from the failed run. When the fix changes package contents, omit those reuse inputs so the workflow packs a new tarball. Live-only targeted reruns skip the E2E images and build only the live-test image. Release-path normal mode fans out into smaller Docker chunk jobs:

• core
• package-update-openai
• package-update-anthropic
• package-update-core
• plugins-runtime-plugins
• plugins-runtime-services
• plugins-runtime-install-a
• plugins-runtime-install-b
• plugins-runtime-install-c
• plugins-runtime-install-d
• bundled-channels

OpenWebUI is folded into plugins-runtime-services for full release-path coverage and keeps a standalone openwebui chunk only for OpenWebUI-only dispatches. The legacy package-update, plugins-runtime-core, plugins-runtime, and plugins-integrations chunks still work as aggregate aliases for manual reruns, but the release workflow uses the split chunks so provider installer checks, plugin runtime checks, bundled plugin install/uninstall shards, and bundled-channel checks can run on separate machines. The bundled-channel runtime-dependency coverage inside bundled-channels uses the split bundled-channel-* and bundled-channel-update-* lanes rather than the serial bundled-channel-deps lane, so failures produce cheap targeted reruns for the exact channel/update scenario. The bundled plugin install/uninstall sweep is also split into bundled-plugin-install-uninstall-0 through bundled-plugin-install-uninstall-7; selecting the legacy bundled-plugin-install-uninstall lane expands to all eight shards.

Package Acceptance

Use the manual Package Acceptance workflow when the question is "does this installable package work as a product?" rather than "does this source diff pass Vitest?"

In release validation, treat Package Acceptance as the package-candidate shard inside the larger release umbrella, not as a competing full-test path. Full Release Validation and private release gauntlets should call Package Acceptance for tarball resolution, Docker product/package proof, and optional Telegram QA against the same resolved package-under-test artifact; keep orchestration, secret policy, blocking/advisory status, and evidence rollup in the caller.

Good defaults:

gh workflow run package-acceptance.yml --ref main \
  -f source=npm \
  -f workflow_ref=main \
  -f package_spec=openclaw@beta \
  -f suite_profile=product \
  -f telegram_mode=mock-openai

Npm candidate selection:

• Resolve the registry immediately before dispatch: npm view openclaw dist-tags --json --prefer-online --cache /tmp/openclaw-npm-cache-verify-$$ and npm view openclaw@beta version dist.tarball dist.integrity --json --prefer-online --cache /tmp/openclaw-npm-cache-verify-$$.
• If Peter asks for "latest beta", use source=npm with package_spec=openclaw@beta, then record the resolved version from npm view or the workflow summary.
• For reruns, release proof, or comparing one known package, prefer the exact immutable spec: package_spec=openclaw@YYYY.M.D-beta.N or package_spec=openclaw@YYYY.M.D.
• For stable package proof, use package_spec=openclaw@latest only when the question is explicitly the current stable dist-tag; otherwise pin the exact version.
• source=npm only accepts registry specs for openclaw@beta, openclaw@latest, or exact OpenClaw release versions. Do not pass semver ranges, git refs, file paths, tarball URLs, or plugin package names there.
• If the candidate is a tarball URL, use source=url with package_sha256. If it is an Actions tarball artifact, use source=artifact. If it is an unpublished source candidate, use source=ref with a trusted ref or SHA.
• Package acceptance tests exactly the selected package candidate. Do not apply openclaw update --channel beta fallback semantics here; if beta is absent, stale, older than latest, or points at a broken tarball, report that tag state instead of silently testing latest.

Profiles:

• smoke: quick confidence that the tarball installs, can onboard a channel, can run an agent turn, and basic gateway/config lanes work.
• package: release-package contract. Adds installer/update, doctor install switching, bundled plugin runtime deps, plugin install/update, and package repair lanes. This is the default native replacement for most Parallels package/update coverage.
• product: package profile plus broader product surfaces: MCP channels, cron/subagent cleanup, OpenAI web search, and OpenWebUI.
• full: split Docker release-path chunks with OpenWebUI.
• custom: exact docker_lanes list for a focused rerun.

Candidate sources:

• source=npm: openclaw@beta, openclaw@latest, or an exact release version.
• source=ref: pack package_ref using the trusted workflow_ref harness. This intentionally separates old package commits from new workflow/test code.
• source=url: HTTPS .tgz plus required package_sha256.
• source=artifact: download one .tgz from artifact_run_id/artifact_name.

Ref model:

• gh workflow run ... --ref <workflow-ref> selects the workflow file revision GitHub executes.
• workflow_ref is the trusted harness/script ref passed to reusable Docker E2E.
• package_ref is the source ref to build when source=ref. It can be an older branch/tag/SHA as long as it is reachable from an OpenClaw branch or release tag.

Example: run latest package acceptance harness against an older trusted commit:

gh workflow run package-acceptance.yml --ref main \
  -f workflow_ref=main \
  -f source=ref \
  -f package_ref=<branch-or-sha> \
  -f suite_profile=package \
  -f telegram_mode=mock-openai

Use telegram_mode=mock-openai or telegram_mode=live-frontier when the same resolved package-under-test tarball should also run through the Telegram QA workflow in the qa-live-shared environment. The standalone Telegram workflow still accepts a published npm spec for post-publish checks, but Package Acceptance passes the resolved artifact for source=npm, ref, url, and artifact. Use telegram_mode=none only when intentionally skipping Telegram credentialed package proof for a focused rerun.

Docker E2E images never copy repo sources as the app under test: the bare image is a Node/Git runner, and the functional image installs the same prebuilt npm tarball that bare lanes mount. scripts/package-openclaw-for-docker.mjs is the single packer for local scripts and CI and validates the tarball inventory before Docker consumes it. scripts/test-docker-all.mjs --plan-json is the scheduler-owned CI plan for image kind, package, live image, lane, and credential needs. Docker lane definitions live in the single scenario catalog scripts/lib/docker-e2e-scenarios.mjs; planner logic lives in scripts/lib/docker-e2e-plan.mjs. scripts/docker-e2e.mjs converts plan and summary JSON into GitHub outputs and step summaries. Every scheduler run writes .artifacts/docker-tests/**/summary.json plus failures.json. Read those before rerunning. Lane entries include command, rerunCommand, status, timing, timeout state, image kind, and log file path. The summary also includes top-level phase timings for preflight, image build, package prep, lane pools, and cleanup. Use pnpm test:docker:timings <summary.json> to rank slow lanes and phases before deciding whether a broader rerun is justified.

Skill install proof: use pnpm test:docker:skill-install or targeted docker_lanes=skill-install for live ClawHub skill-install validation. The lane installs the package tarball in a bare runner, keeps skills.install.allowUploadedArchives=false, resolves the current live slug from openclaw skills search, installs it, and verifies .clawhub origin/lock metadata. Prefer this checked-in script over inline heredoc Testbox recipes.

Cheap Docker Reruns

First derive the smallest rerun command from artifacts:

pnpm test:docker:rerun <github-run-id>
pnpm test:docker:rerun .artifacts/docker-tests/<run>/failures.json

The script downloads Docker E2E artifacts for a GitHub run, reads summary.json/failures.json, and prints a combined targeted workflow command plus per-lane commands. Prefer the combined targeted command when several lanes failed for the same patch:

gh workflow run openclaw-live-and-e2e-checks-reusable.yml \
  -f ref=<sha> \
  -f include_repo_e2e=false \
  -f include_release_path_suites=false \
  -f include_openwebui=false \
  -f docker_lanes='install-e2e bundled-channel-update-acpx' \
  -f include_live_suites=false \
  -f live_models_only=false

That path still runs the prepare job, so it creates a new tarball for <sha>. If the SHA-tagged GHCR bare/functional image already exists, CI skips rebuilding that image and only uploads the fresh package artifact before the targeted lane job. Do not rerun the full release path unless the failed lane list or touched surface really requires it.

Docker Expected Timings

Treat these as ballpark. Blacksmith queue time, GHCR pull speed, provider latency, npm cache state, and Docker daemon health can dominate.

Current local timing artifact (.artifacts/docker-tests/lane-timings.json) has these rough bands:

• Tiny lanes, seconds to under 1 minute: agents-delete-shared-workspace ~3s, plugin-update ~7s, config-reload ~14s, pi-bundle-mcp-tools ~15s, onboard ~18s, session-runtime-context ~20s, gateway-network ~34s, qr ~44s.
• Medium deterministic lanes, ~1-5 minutes: npm-onboard-channel-agent ~96s, openai-image-auth ~99s, bundled channel/update lanes usually ~90-300s when split, openwebui ~225s, mcp-channels ~274s.
• Heavy deterministic lanes, ~6-10 minutes: bundled-channel-root-owned ~429s, bundled-channel-setup-entry ~420s, bundled-channel-load-failure ~383s, cron-mcp-cleanup ~567s.
• Live provider lanes, often ~15-20 minutes: live-gateway ~958s, live-models ~1054s.
• Installer/release lanes: install-e2e and package-update paths can vary widely with npm, provider, and package registry behavior. Budget tens of minutes; prefer GitHub targeted reruns over local repeats.

Default fallback lane timeout is 120 minutes. A timeout usually means debug the lane log/artifacts first, not “run the whole thing again.”

Failure Workflow

1. Identify exact failing job, SHA, lane, and artifact path.
2. Read failures.json, summary.json, and the failed lane log tail.
3. Use pnpm test:docker:rerun <run-id|failures.json> to generate targeted GitHub rerun commands.
4. If the lane has rerunCommand, use that only as a local starting point.
5. For Docker release failures, dispatch targeted docker_lanes=<failed-lane> on GitHub before considering local Docker.
6. Patch narrowly, then rerun the failed file/lane only.
7. Broaden to pnpm check:changed or CI only after the isolated proof passes.

When To Escalate

• Public SDK/plugin contract changes: run changed gate plus relevant extension validation.
• Build output, lazy imports, package boundaries, or published surfaces: include pnpm build.
• Workflow edits: run pnpm check:workflows.
• Release branch or tag validation: use release docs and GitHub workflows; avoid local Docker unless Peter explicitly asks.