# Preserve native Telegram answers and use compatible free inference routes

Status: implemented and verified. The user owns publication and runtime replacement.

## Anchored incident

Telegram input 3363 (a backtick followed by “hey”), update activity
`telegram:1e8af20a7e273e82`, finalization
`telegram-64ac9937dc7647a8-1791853421483-1` ran on 1.0.846, PID 2409421,
started 2026-10-13T00:32:14.464Z. Its bundle SHA256 matches the previous
verified artifact:
`f568a8a54e037eb22f21193f7f589f8769b6b94369c0e9dfa7446ca6e8a5f87b`.
This incident is not explained by an older installed build.

Turn 0 returned ordinary assistant text: “Hey! How can I assist you today?”
The transcript and turn-1 context retain that answer. Native Telegram
disabled text completion, so the runner demanded another generation to
call task_complete. Turn 1 exhausted the remote routes and the usable
answer never acquired terminal delivery authority. The Activity panel
obscured the successful first response.

The public registry at https://robit-man.github.io/inference/ also exposed
two reproducible routing gaps: discovery lost LLM7's provider-default
selector and chose the first catalog entry, and preferred-model
incompatibility skipped a provider with other compatible models. The
catalog-less ch.at chat endpoint was excluded even from tool-free requests.
Asynchronous jobs, disabled relays, and non-chat services require distinct
adapters; their presence in the page does not establish native tool support.

## Ordered work

1. Admit an exact current native assistant candidate through host semantic
   assessment and real effect checks. Avoid a redundant task_complete
   generation. Preserve run, epoch, turn, owner, steering, candidate and
   physical evidence fences, terminal receipts and the actual Telegram ACK.
   No executable completion tool call may be synthesized from prose.
2. Continue actual work when a promise lacks required mutation, observation,
   verification or delivery. An unavailable auxiliary assessment may permit
   delivery of current useful text under existing unavailable-review policy;
   it must leave the native goal unresolved and never forge fulfillment.
3. Preserve known selector and transport hints while merging discovered
   provider capabilities. Enumerate bounded, compatible catalog alternatives
   deterministically, deduplicate actual routes, preserve whole-input capacity
   checks, deadlines, manual approval, Stop and provider-specific credentials.
4. Support catalog-less ch.at only for tool-free chat using its documented
   omitted-model request. Do not invent a model identifier on the wire or
   advertise tools, images, or unsupported job protocols.
5. Test ordinary native text through the actual backend/runner/bridge path,
   including unavailable assessment, real missing effects, stale ownership,
   cancellation and provider ACK failure. Exercise registry selector and
   catalog alternative regressions without live inference or Telegram sends.
6. Rebuild the workspaces and actual CLI artifact, verify bundle markers and
   package gates, then commit and push only task-owned changes. Preserve
   unrelated dirty work and user-owned release versions. The user owns npm
   publication and runtime replacement.

No service restart, local model request, CUDA workload change or external
Telegram message is part of this repair's verification.

## Result and verification

Native private and public replies now submit exact current plain text for
host review and terminal delivery custody without another completion
generation. Rejected candidates carry bounded missing-effect feedback into
the next primary request. Stop, steering reconciliation, stale candidate
rejection, ordinary prose-loop limits and actual provider acknowledgements
remain enforced. The host neither executes a synthetic task_complete nor
marks an unknown native goal fulfilled when auxiliary inference is unavailable.
Selected/generated artifacts retain their explicit delivery boundary.

The runtime merges known provider selector/auth/transport hints with page
capabilities, tries bounded compatible catalog alternatives after giving
providers their first attempt, and reads nested context/tool metadata.
Explicit non-chat, moderation and paid-only catalog entries are excluded.
ch.at's text-only route omits the model on the wire and needs no catalog;
onboarding also supports it while retaining a synthetic readiness probe.

Verification:

- 136 registry, routing and existing fallback checks passed.
- 18 host-admission/model-wire checks and 69 existing completion/steering
  authority checks passed.
- 12 new actual backend/runner/Telegram bridge integration cases passed with
  mocked network responses, including public topic-bound ACK, unavailable
  assessment, real file mutation, review feedback, Stop and stale ownership.
- 89 existing Telegram/fallback checks, 11 header route checks and 61
  onboarding checks passed. Total: 396 regression cases.
- Two live synthetic requests through the production backend succeeded:
  streaming text via ch.at in 4821 ms, and streaming native function calling
  via Hermes in 1764 ms. Both recorded zero local dispatches and executed no
  generated action. These probes contained only synthetic public test text.
- Full workspace rebuild and staging audit passed. The tarball audit passed
  with 531 files and zero warnings. Fresh isolated global npm installation
  with lifecycle scripts enabled and compilers blocked passed, including
  SQLite, CLI/library, companion install and PATH checks.
- All 12 mocked transport scenarios passed against the exact installed
  tarball bundle. Installed and staged runtime SHA256 match:
  `f5581b26f6b472a9af893e906670aabb393bc4c63944d2daeadbec245809eb01`.

Logs:
`/tmp/omnius-keyless-routing-tests-20261012.log`,
`/tmp/omnius-host-text-admission-focused.log`,
`/tmp/omnius-host-text-admission-regressions.log`,
`/tmp/omnius-native-plain-text-admission-20261012.log`,
`/tmp/omnius-plain-answer-cli-regressions-20261012.log`,
`/tmp/omnius-plain-answer-route-header-20261012.log`,
`/tmp/omnius-onboarding-catalogless-tests-20261012.log`,
`/tmp/omnius-plain-answer-live-remote-text-20261012.json`,
`/tmp/omnius-plain-answer-live-remote-tools-20261012.json`,
`/tmp/omnius-plain-answer-publish-build-final-20261012.log`,
`/tmp/omnius-plain-answer-tarball-audit-20261012.log`,
`/tmp/omnius-plain-answer-fresh-install-20261012.log` and
`/tmp/omnius-plain-answer-installed-verification-20261012.json`.

The already running 1.0.846 bot still uses the earlier f568a8a5 bundle.
The next user-owned patch publication/update and restart must load this repair;
the isolated package retained the user's staging version without publishing it.
