Zoneasetech Aiworker Aiworker Soul E2e Sampling

AIWorker Soul E2E Sampling

Use this skill to operate the AIWorker official Soul output-quality sampling loop. This is an execution workflow for real sampling evidence, not a generic AIWorker development guide and not a skill-authoring template.

Trigger Gate

Before using this skill, confirm all three gates:

1. AIWorker repo identity: Locate the git root from the current working directory. Confirm root AGENTS.md contains AIWorker Agent Bootstrap, the five canonical docs exist, scripts/e2e-soul-sampling.ts exists, and official Soul packages exist under souls/*.
2. Sampling operation intent: The user asks to run, continue, triage, retest, or tune the official Soul output-quality sampling loop.
3. Official Soul or evidence target: The request targets existing official Soul AGENTS.md, projected skills, or evidence under tmp/e2e-soul-sampling/.

If any gate fails, do not use this skill. For ambiguous prompts such as "跑 e2e" or "改 product-manager skill", ask one clarifying question to separate the Soul sampling loop from ordinary project development.

Do not use this skill for ordinary AIWorker feature work, generic E2E/contract/smoke tests, normal Soul authoring, new skill creation, sampling harness development, docs-only architecture work, or AGENTS/SKILL edits not driven by sampling evidence.

Zero-Trust Start

After the trigger gate passes, re-read current repo truth before choosing work:

• AGENTS.md
• docs/architecture.md
• docs/protocol.md
• docs/runtime.md
• docs/soul-authoring.md
• docs/testing.md
• scripts/e2e-soul-sampling.ts
• scripts/e2e-soul-sampling.test.ts
• git status --short
• relevant recent commits

For execution tasks, run early:

bun run docs:check
bun run test:contracts

If test:contracts fails because a local dependency is missing, such as Cannot find package 'zod', try bun install and rerun the gate before declaring architecture drift.

Real Engine Rule

Unit or static tests for scripts/e2e-soul-sampling.ts may mock CLI behavior. E2E sampling must use real Worker CLI and real Codex invocations. Do not replace LLM-engine behavior with fake outputs, golden text, dry-run evidence, or mock-only validation.

Canonical real-run commands:

AIWORKER_E2E_RUN_ID=full-freeform AIWORKER_E2E_REASONING=high bun scripts/e2e-soul-sampling.ts run --soul aiworker-freeform
AIWORKER_E2E_RUN_ID=full-google-ads AIWORKER_E2E_REASONING=high bun scripts/e2e-soul-sampling.ts run --soul google-ads
AIWORKER_E2E_RUN_ID=full-hr-manager AIWORKER_E2E_REASONING=high bun scripts/e2e-soul-sampling.ts run --soul hr-manager
AIWORKER_E2E_RUN_ID=full-product-manager AIWORKER_E2E_REASONING=high bun scripts/e2e-soul-sampling.ts run --soul product-manager
AIWORKER_E2E_RUN_ID=full-software-support AIWORKER_E2E_REASONING=high bun scripts/e2e-soul-sampling.ts run --soul software-support

Use unique run IDs for retests, for example full-product-manager-agentsfix. Prefer serial runs unless the user explicitly wants parallel sampling and the machine can support it.

Evidence Review

Evidence lives under:

tmp/e2e-soul-sampling/<runId>/
  manifest.json
  scorecards/*.json
  events/*.json

scorecard.status=pass means the invocation completed; it does not prove output quality. Always read assistant event text before judging quality:

jq -r '[.events[] | (.payloadJson.data.text // empty)] | join("")' \
  tmp/e2e-soul-sampling/<runId>/events/<caseId>.json

Write ignored working summaries such as tmp/e2e-soul-sampling/<runId>/findings.md when reviewing multiple cases. Do not commit tmp/ evidence.

Finding Routing

Classify before editing:

Edit the smallest asset that explains the failure. Shared knowledge or template changes require repeated evidence across multiple cases.

Fix And Retest Loop

For each verified finding:

1. Classify the failure from assistant event text.
2. Edit the smallest responsible Soul asset.
3. Build or validate the affected Soul package.
4. Confirm tracked dist/engine-assets output is synchronized when the package generates it.
5. Retest with real Codex using the affected Soul or the narrowest supported case.
6. Re-read assistant event text and update findings.
7. Commit the verified slice without tmp/ evidence.

Example validation commands:

bun run --filter '@zonease/aiworker-product-manager' validate
bun run --filter '@zonease/aiworker-software-support' validate

Completion Gate

Before claiming completion, run the smallest fresh verification set that proves the touched surface. Common commands:

bun run docs:check
bun run test:contracts
bun test scripts/e2e-soul-sampling.test.ts --timeout=30000
bun test packages/worker-runtime/src/worker/executor.test.ts --timeout=30000

For code changes, run code-review-graph unless the change is docs-only, instruction-only, or pure formatting.

Source: ZonEaseTech/aiworker — distributed by TomeVault.