Skip to content

harness-output.md

Latest commit

 

History

History
92 lines (60 loc) · 3.28 KB

harness-output.md

File metadata and controls

92 lines (60 loc) · 3.28 KB
Raw

Harness Output

中文 | English

OpenCode++ has two output layers so normal coding work stays readable without losing audit detail.

Compact status

prepare, retrieve, evaluate, and next return structured JSON. Their humanReadable field is compact by default:

OpenCode++ ✓ Verified

Changed
3 files

Checks
✓ npm test
✓ npm run typecheck

Status
Ready to finalize

When a current check fails, the result is explicit:

OpenCode++ ✗ Repair required

Failed
npm test

Next
Fix test/auth/session.test.ts.

When OpenCode++ cannot prove a source change with an executable verifier:

OpenCode++ ⚠ Human review

Need you
No executable verification command was found for source changes.

Suggested
Configure a test command or manually review the diff.

Only non-empty high-signal action categories may appear in compact text. The result never describes a recommended command as a passed check: an unexecuted command is shown as a suggestion or next action.

Structured action summary

The JSON actionSummary remains the compatibility and audit interface. It keeps all categories even when a category is empty:

  • observed: a file, finding, state, or context signal was recorded;
  • prevented: a boundary, command, or policy risk was stopped;
  • requested: an action is required before the loop can proceed;
  • repaired: an intervention changed state but is not verified;
  • verified: fresh command or CI evidence matches the current working tree;
  • unresolved: a blocker remains active.

The structured visualization, interventions, findings, missingEvidence, and requiredCommands fields are also retained. They are the source of truth for artifacts and API clients.

Detailed Dashboard

Use opencode_plusplus_dashboard when you want the full view. It includes:

  • phase progress from Plan through Finalize;
  • selected and rejected files and rejection reasons;
  • boundaries, findings, missing evidence, and required commands;
  • working-tree hash and evidence freshness;
  • intervention counts and decision basis;
  • final decision and next action.

Human-review results also expose a separate dashboard field so a compact status can remain readable while the complete recorded facts are available for review. The snapshot is persisted at .agent-context/sidecar/visualization.json; the latest human report is at .agent-context/sidecar/latest.md.

The Dashboard shows recorded facts and deterministic decision inputs. It does not expose hidden model chain-of-thought, and it does not add a second model call.

Toast transitions

Toasts are state-change notifications, not a log of every finding. OpenCode++ emits at most one deduplicated toast for each transition:

  • verification started;
  • repair required;
  • human review required;
  • verified.

The same decision and evidence state does not repeatedly notify the user. Detailed events remain in the sidecar trace and intervention ledger.

What counts as verified

✓ means the Harness found valid command or CI evidence captured against the current working-tree hash. It does not mean the model's explanation, a manual claim, an old test run, a Context Pack, or an annotation is proof. A successful command is evidence for the gate; business correctness may still require human review.