# WO-01: Keep internal diagnostics out of progress and classify browser effects truthfully

Status: complete — integrated and verified offline

## Anchors

- `packages/orchestrator/src/agenticRunner.ts`: runtime guidance, claim discovery,
  tool deferral, `_visualToolEffect`, evidence-result processing.
- Existing `AgenticEvent.visibility` and TUI/Telegram telemetry filtering.

## Acceptance

- [x] Runtime/adversary guidance and routine discovery/deferral bookkeeping are
  telemetry; retain actual user progress and actionable errors.
- [x] A successful DOM evaluation returning `[]` does not assert browser mutation
  or queue missing-mutation-proof guidance.
- [x] Genuine interaction evidence and failure handling continue to work.
- [x] Diagnostic archival is preserved; ordinary sub-agent views stay quiet.
- [x] Focused producer/consumer regression tests pass; scoped implementation committed.

## Implementation and verification

Runtime guidance, model routing, claim discovery, and tool-deferral receipts now
carry `visibility: "telemetry"`. Existing TUI and Telegram consumers hide them
from normal progress while preserving the diagnostic archive. Genuine runtime
guidance still requests internal review, and observed verification failures stay
visible.

Follow-up: expiry, verified resolution, and supersession messages from the failure
lifecycle also use telemetry. Retiring stale records does not remove current
failures; real tool failures, capacity rejection, and verification results keep
their existing visibility. Consumer regressions cover all four lifecycle receipt
variants in the cold archive, normal Activity, and TUI child views.

Browser evaluation has an unknown effect unless structured evidence establishes
more. A returned value, including `"[]"`, no longer creates a false mutation
critique or its trajectory-review trigger. Supplied contradicted/proven evidence
still follows the existing verification path. Known interactions without proof
still request observation, with accurate wording that the action may change state.
This change does not claim arbitrary JavaScript is read-only or turn its result
into completion proof.

Initial verification: 22 focused checks passed across orchestrator
`activity-diagnostic-visibility.test.ts` (15),
`completion-provenance-visual-evidence.test.ts` (3),
`model-resolution-telemetry.test.ts` (1), and CLI
`activity-diagnostic-visibility.test.ts` (3). Tests use isolated files and mocked
inference; no running session or external service was changed. `git diff --check`
passed. Implementation commit subject: `fix: keep runtime diagnostics out of progress`.

Integration found and fixed two Telegram consumers that wrote status content
directly into sub-agent panes: the primary runner callback and the native child
callback. Both now honor telemetry visibility; the primary typed event/archive
path still retains the diagnostics. Production callback regressions, the expanded
15-case CLI visibility/archive matrix, and the integrated 55-case orchestrator
set pass. See WO-02 and WO-05 for delivery and complete verification evidence.

Failure-lifecycle follow-up verification: 41 checks passed across orchestrator
`activity-diagnostic-visibility.test.ts` (16), `run-observability.test.ts` (10),
and CLI `activity-diagnostic-visibility.test.ts` (15). These cover expiry without
losing active failure evidence, all lifecycle receipt variants in ordinary and
archival views, and unchanged genuine post-action audit/completion behavior.
`git diff --check` passed. Follow-up commit subject:
`fix: classify failure lifecycle bookkeeping as telemetry`.
