# Telegram reply presentation when an auxiliary review is unavailable

Status: implemented and verified. Publication and runtime replacement remain
with the user.

## Live incident

The user's 17:08 PDT `quick ping!` received `pong`, followed by an incorrect
generic assertion that the task remained incomplete. This was the repaired
1.0.845 runtime, PID 2042875 started at 17:05:34 PDT, not stale code. Its
bundle SHA256 was
`a9cbdc655c3f3be3c254f09caaa9149c8677cb9858cb12940142e24dc9eb514e`.

Input 3360, update 666870213, runner `telegram:a0e3cbd0c36f1126`, finalization
`telegram-64ac9937dc7647a8-1791850093839-1` and conversation event line 1905
anchor the trace. The final text was acknowledged at 17:08:48 PDT. The
auxiliary native goal assessment was `unavailable`, not `incomplete`.
User-visible telemetry recorded local HTTP 503, LLM7 HTTP 429 and OVH empty
response. Detailed review errors were emitted to the live TUI, not retained
in the finalization's generic unavailable assessment.

The runner's earlier `incomplete_verification` receipt was awaiting the
visible-response provider acknowledgement. It did not demonstrate unfinished
work. The bridge appended the failure sentence for every assessment that
was not ready, conflating review transport failure with a negative verdict.

## Work orders

1. Exclude only unavailable auxiliary reviews from the generic incomplete
   suffix when the current runner has not declared a partial answer. Keep
   actual partial, incomplete and uncertain verdict behavior intact.
2. Preserve current owner, input, epoch, steering, exact candidate and
   physical evidence checks before sending useful text despite an unavailable
   review. No failed or stale review grants completion or delivery authority.
3. Preserve the unavailable assessment, retained goal and real receipts;
   do not forge fulfillment, consume unresolved work, or automatically replay
   object actions after the provider acknowledges the visible reply.
4. Exercise real runner-to-delivery fixtures for ping, transport failure,
   empty/truncated assessment, useful completed read work, genuine partial
   work and stale owner/candidate/steering bindings. Include public native
   replies as well as admin DMs.
5. Rebuild all workspace packages, audit the actual artifact and deliver a
   scoped verified commit to the tracked branch. Preserve unrelated bridge
   edits and user-owned release versions. Publication remains with the user.

No greeting classifier, model assertion of external effects, live inference,
Telegram test sends, service restart or CUDA workload change is required.

## Result and verification

Unavailable reviews no longer append the incomplete-task sentence to useful
current replies. Genuine partial answers and negative/uncertain verdicts
retain their existing reporting. The unavailable review and retained goal
remain unresolved internally; no new completion authority or automatic
object-work replay is created. Delivery additionally checks the unavailable
assessment's exact candidate, steering and physical evidence bindings.

- All 62 tests across private native closeout, private runner and public
  runner suites passed. Cases include exact `pong` after HTTP 503, empty or
  truncated auxiliary output, useful completed read work, genuine partial
  file work, failed provider acknowledgement and stale owner/epoch/generation,
  candidate and steering bindings.
- Broader native core, steering, initial-instruction, update adoption,
  remote fallback and delivery checks passed 112 of 114 cases. The two panel
  label assertions at update-adoption test lines 234 and 300 also fail on
  clean prior commit `14db5c33`, before completion runs. Their test and panel
  logic are unchanged; they are recorded as existing failures, not relaxed.
- All four bundle-policy tests passed, including rejection of the earlier
  fallback bundle that lacked this unavailable-review repair.
- The full workspace rebuild, bundle audit and 530-file tarball audit passed.
  The exact global tarball installation passed with lifecycle scripts enabled
  and compilers blocked, including SQLite, CLI version, library exports,
  AIWG companion and PATH guidance checks.
- All 12 mocked inference scenarios passed against the actual installed
  bundle. Runtime bundle SHA256:
  `f568a8a54e037eb22f21193f7f589f8769b6b94369c0e9dfa7446ca6e8a5f87b`.

Evidence logs: `/tmp/omnius-native-goal-review-closeout-tests-20261012.log`,
`/tmp/omnius-unavailable-review-integration.log`,
`/tmp/omnius-native-update-adoption-baseline-14db.log`,
`/tmp/omnius-unavailable-review-bundle-policy-tests.log`,
`/tmp/omnius-unavailable-review-publish-build.log`,
`/tmp/omnius-unavailable-review-tarball-audit.log`,
`/tmp/omnius-unavailable-review-fresh-install.log` and
`/tmp/omnius-unavailable-review-installed-verification.json`.

The live bot's earlier 1.0.845 process has not been replaced by these tests.
Its next user-owned publish/update and restart must load the rebuilt code.
