# Telegram outage fallback and user-selected approval mode

Status: implementation complete; distributable verification recorded below.
Publication and runtime installation are reserved for the user.

## Incident anchor

The 2026-10-12 15:06 PDT private Telegram greeting received four identical
incomplete-work reports. The saved activity shows an Ollama native HTTP 503:
`NVIDIA driver recovery is required before new GPU work`. Each attempt ended
at turn zero with no generated tokens. Three same-input continuations followed
the initial failure at 22:06:19, 22:06:24, 22:06:30 and 22:06:39 UTC.

The installed 1.0.841 bundle lacked the source fallback implementation despite
its recent build timestamp. Source-only backend tests did not verify the
deployed Telegram path. Ordinary terminal goal review also pinned its local
endpoint, and fatal backend failures did not retain their no-replay boundary
through Telegram's scheduled input recovery.

## Scoped work orders

1. Exercise actual `processTelegramMessageWork` intake, `AgenticRunner`,
   `OllamaAgenticBackend`, final native goal assessment and provider ACK. The
   reported GPU-recovery 503 must produce one useful greeting through compatible
   free inference, with consumed intake and no queued duplicate.
2. Allow ordinary text goal/source review to use the shared outage cascade.
   Preserve pinned hardware/model and image probes.
3. Persist scope-bound fatal inference receipts. Exhaustion, declined approval
   and interrupted streams must stop same-input recovery, preserve the task,
   and produce one useful diagnosis rather than a generic incomplete report.
4. Add `/fallback auto` (default), `/fallback manual`, and `/fallback status` in
   TUI and admin Telegram. Persist global mode atomically while preserving
   configured provider references and other configuration. Models cannot change
   this host policy. Existing backends load the selected policy per request.
5. Manual mode must wait before any remote discovery or context transmission.
   Telegram approval binds original human, chat, topic, question message and
   work controller. It covers only that input and its final review. TUI uses
   explicit selection. Decline, Stop, missing handler and stale ownership fail
   closed; no elapsed timeout or default selection grants approval.
6. Treat empty responses and clean metadata-only streams as unavailable, while
   preserving valid tool-only results and never replaying started failed streams.
   Recognize broker queue shedding, drain/readiness and wrapped transport errors.
7. Guard the distributable against missing fallback code and exercise its actual
   bundled backend with deterministic HTTP/DNS. Rebuild all workspace packages
   and record package verification separately from deployment. Commit and push
   only these repair files; preserve the original dirty checkout.

## Verification

- Workspace build passed for all 11 projects.
- Orchestrator: 310 tests passed across 13 suites, including 78 fallback-helper
  and 62 policy tests, durable terminal receipts, context recovery, interruption
  and pinned endpoints.
- CLI: 338 tests passed across 19 suites; 2 pre-existing cases were skipped.
  The 6 actual Telegram-native regressions cover free inference after the exact
  503, one diagnosis on exhaustion, manual approval before any HTTPS traffic,
  decline, Stop and stale buttons. Existing native runner, source review,
  onboarding, command and daemon transport tests also passed.
- An additional review-exhaustion suite has 11 existing authentication-fixture
  failures. Running it against an isolated unchanged HEAD bridge reproduced
  all 11 failures before the changed code. No production authentication was
  relaxed and no unrelated fixture redesign was included.
- Packaged backend verification uses
  `node scripts/verify-inference-fallback-bundle.mjs publish/dist/index.js`.
  All 8 checks passed. It checks useful keyless output after mocked local GPU-recovery 503 for unary
  and streaming calls, automatic mode, explicit approval, denial and absent
  approval. It sends no actual inference or Telegram traffic.
- The 20 existing artifact-policy tests and 2 new fallback bundle-policy tests
  passed. Staging dependencies were installed with lifecycle scripts disabled
  to verify the bundle against its declared dependency graph. The verification
  receipt includes the bundle's SHA-256 digest.
- The packed tarball audit passed for 524 files with no warnings. Its backend
  bytes match the SHA-256 digest of the verified bundle; README filename and
  embedded README checks passed. No source maps, tests or stale build caches
  were shipped.

Build and publish from
`/home/roko/Documents/Projects/Adjacent/omnius/omnius-rapid-computer-use-20261012`.
Publish only from that checkout's `publish/` directory. The staged manifest
retains the repository's 1.0.790 version; apply the normal patch bump above the
current npm version before publication.
The unrelated original checkout remains dirty and on an older source revision.
No npm publication, global installation or active-process restart was performed.
