fix(recovery): silence internal assessment warnings by default and sanitize retained notices #286

Merged

karsten merged 2 commits from fix/quiet-recovery-diagnostics into main

2026-11-06 07:19:47 +00:00

karsten commented

2026-11-06 07:19:30 +00:00

Owner

Problem

Internal completion/recovery inference failures were shown as alarming German operation failures. Exhaustion bypassed recoveryDiagnostics:false and repeated complete user requests and up to 8000 response characters. Retained queued notices could survive the upgrade.

Fix

All recovery assessment reasons, including 3-attempt exhaustion, respect the resolved channel diagnostic setting. Optional notices are concise English text with bounded identifiers/diagnostic enums and an inspect command, no task text/reply/provider output/reset instructions. Retry/termination/expiry/stop/admission rules are unchanged.

Restore migrates only reserved host recovery IDs with a known diagnostic prefix: default-off and stale terminal tasks are silent; opted-in pending notices are sanitized with the same delivery ID. Migration saves before any delivery. Ordinary replies, quoted warning text, actual blockers and approvals are not filtered. Existing delivered Matrix messages are not deleted.

Verification

Typecheck passed. Complete final suite 1551/1551 passed. Targeted tests cover opt-in/default-off/per-channel opt-out, exhausted budgets, no fourth assessment or worker replay, durable reset path, actual approvals, stop/deadline fencing, disk failure during queued-notice migration, restart/deduplication, unchanged retained task/reply and preservation of real replies/approvals.

Delivery

Deploy the normal shared image and native releases to the existing agents and local installation. Separate model-protocol/reload and foreground restart issues are tracked independently; this PR does not claim those fixes.

## Problem Internal completion/recovery inference failures were shown as alarming German operation failures. Exhaustion bypassed recoveryDiagnostics:false and repeated complete user requests and up to 8000 response characters. Retained queued notices could survive the upgrade. ## Fix All recovery assessment reasons, including 3-attempt exhaustion, respect the resolved channel diagnostic setting. Optional notices are concise English text with bounded identifiers/diagnostic enums and an inspect command, no task text/reply/provider output/reset instructions. Retry/termination/expiry/stop/admission rules are unchanged. Restore migrates only reserved host recovery IDs with a known diagnostic prefix: default-off and stale terminal tasks are silent; opted-in pending notices are sanitized with the same delivery ID. Migration saves before any delivery. Ordinary replies, quoted warning text, actual blockers and approvals are not filtered. Existing delivered Matrix messages are not deleted. ## Verification Typecheck passed. Complete final suite **1551/1551 passed**. Targeted tests cover opt-in/default-off/per-channel opt-out, exhausted budgets, no fourth assessment or worker replay, durable reset path, actual approvals, stop/deadline fencing, disk failure during queued-notice migration, restart/deduplication, unchanged retained task/reply and preservation of real replies/approvals. ## Delivery Deploy the normal shared image and native releases to the existing agents and local installation. Separate model-protocol/reload and foreground restart issues are tracked independently; this PR does not claim those fixes.

karsten added 2 commits

2026-11-06 07:19:30 +00:00