Harness Engineering Compound

Progressive-disclosure entrypoint for stage orchestration and durable learning capture in Harness Engineering.

Philosophy

• Route to the correct stage before acting.
• Capture verified outcomes as reusable team knowledge.
• Keep orchestration, evidence gathering, and durable learning capture explicit instead of collapsing them into generic advice.

When to use

• Route/resume work from the correct Harness Engineering stage.
• Capture a verified fix in docs/solutions/ and avoid duplicate docs by refreshing existing artifacts when overlap is high.
• Decide whether a request needs lifecycle orchestration, solved-problem capture, or a resume-from-stage handoff before downstream work starts.

Inputs

• Request, artifacts, repo context, and linked Linear issues.

Outputs

• schema_version: 1 when structured; result, validation, blockers, and next Harness Engineering action.

Procedure

1. Select lifecycle mode using artifact-first evidence.
2. For lifecycle routing, continue from the earliest incomplete or untrusted stage.
3. If the user intent is ambiguous between stage orchestration and solved-problem capture, ask one direct question before continuing.
4. For learning capture, preserve full as the default and treat compact-safe as explicit opt-in.
5. For learning capture, validate solved evidence, gather supporting inputs, and write exactly one durable solution artifact.
6. If past agent sessions are needed as evidence, route broad session inventory/extraction to skill-refactor; if the result is a reusable workflow, route skill creation to skillify.
7. If helpers are used during learning capture, they return text only; the orchestrator writes the final artifact.
8. If overlap with an existing solution is high, refresh the existing doc instead of creating a duplicate.
9. Recommend he-compound-refresh only when adjacent stale or overlapping docs need selective follow-up beyond the current artifact.

Validation

• Confirm mode selection matches available evidence.
• Confirm stage recommendation names the exact next Harness Engineering command or stage.
• For learning capture, verify solved status before writing docs, keep the one-file-write contract explicit, and ensure duplicate-avoidance logic is explicit.
• Any refresh recommendation must be narrow and evidence-backed rather than a blanket cleanup suggestion.
• Fail fast: stop at first failed gate and do not continue.

Constraints

• Redact secrets, credentials, tokens, and sensitive data by default.
• Do not document unverified fixes as durable solutions.
• Do not treat he-compound as a substitute for implementation, debugging, review, or refresh stages that still need to happen.
• Do not turn generic session-history search into a Harness Engineering stage; consume skill-refactor or skillify outputs when session evidence is relevant.
• Do not let helper agents write intermediate files during learning capture.
• Do not recommend deleting or ignoring protected process artifacts just to simplify the handoff.
• Apply the context-disposition policy: move important still-valid context to references and index it when meaningful; intentionally discard stale, duplicated, unsafe, superseded, or low-signal text.

Anti-patterns

• Guessing stage progression without checking existing artifacts.
• Recording unresolved incidents as solved knowledge.
• Creating a duplicate solution doc when a high-overlap artifact should be refreshed.
• Returning broad "use Harness Engineering" guidance without naming the exact mode and next stage.

Full Context

• Assets: icon-small.png, icon-large.png

Examples

• "Can you inspect this feature's artifacts, resume from the right Harness Engineering stage, and tell me the next exact action?"
• "We fixed the same CI failure twice. Capture the solved workflow once and refresh the existing solution if it overlaps."
• "Help me coordinate brainstorm, spec, plan, work, and review for this Linear issue without skipping the earliest weak stage."