# Primary checkout delivery and remote output capacity

Status: implementation complete and verified in the primary checkout.
Publication remains with the user.

## Incident and evidence

At 2026-10-12 16:12 PDT, Telegram input 3348 received four generic incomplete
reports. The initial run and three scheduled continuations failed at turn zero
with the local broker's NVIDIA recovery HTTP 503. No tool calls ran.

Installed `omnius@1.0.842` and the primary checkout's old publish bundle shared
SHA-256 `844d817565852016316801b291ba0bb8ef2968d5f79b890e75b73ef9dbc263e6`.
Both lacked all six fallback implementation markers. The running Telegram TUI
was PID 1146808; the daemon was PID 1146606. The version increase and restart
did not bring in the source repair from the separate worktree. The primary
checkout was still 59 commits behind at `f1b2d5f9`.

The installed bundle verifier rejects that exact installed artifact before any
inference or provider discovery. Prior deterministic tests exercised the
repaired worktree, not these deployed bytes.

## Durable work orders

1. Reconcile the primary checkout with delivered `1b838693`, preserving every
   unrelated local edit and untracked file. Apply and review a three-way merge
   preview before changing the checkout; retain a private backup and Git stash.
2. Rebuild the primary publication package. Its normal publisher must execute
   the fallback marker guard and pre-pack audit from this checkout.
3. Exercise actual native Telegram intake with a realistic remote model window.
   Verify the complete prompt, tools and response schemas remain intact.
4. Adapt only the remote output ceiling to its advertised context/output limits.
   A large local maximum is an upper bound, not a minimum remote requirement.
   Keep the local request unchanged, preserve endpoint pinning, and reject
   providers whose capacity cannot hold the complete input plus viable output.
5. Verify the actual primary bundled backend, automatic/manual modes, Stop,
   exhaustion diagnosis, no duplicate recovery, and package contents. Commit and
   push only this repair; do not include preserved unrelated edits.

## Reconciliation

The primary checkout now contains the delivered source repair and all preceding
computer-use/progress work. Twenty-five modified tracked files were preserved
through a reviewed merge; all untracked files stayed in place. The four small
conflicts retained both desktop/history hooks, the local status-display filter,
and both diagnostic test changes.

Recovery stash: `140ecbfb7ca255aeeae8760ee7277bd9593623c5`.
Private backup: `/home/roko/.cache/omnius-primary-integration-20261012-hrn6abzp`.
The stash is retained as a safety copy; its edits are already restored.

## Additional production-size defect

The incident's local output allowance was 218,406 tokens. A read-only Hermes
model-catalog request advertised `max_model_len: 65536` for
`turboderp/Qwen3.8-27B-exl3`. No inference was sent to that service.

A new native Telegram regression failed on the prior repaired source: its
complete input estimated 28,568 tokens, local output allowance was 217,847,
and the remote window was 65,536. The old compatibility check made zero remote
generation attempts despite enough room for the complete request and about
33,035 output tokens after the safety margin. The previous fixture's
1,048,576-token remote catalog hid this defect.

Fallback now clones the request with an output ceiling bounded by remaining
remote capacity and any advertised `max_completion_tokens`. Known output
limits remain explicit on the wire. Messages, tool schemas, response format,
ownership and the original local request are preserved. Full input overflow
still excludes a provider.

## Verification and activation

- All 11 workspace projects rebuilt successfully and the staging audit passed.
- Primary checkout: 329 orchestrator tests and 233 CLI tests passed across 26
  suites, including all 7 native Telegram regressions. Two existing cases were
  skipped. Daemon split-stream, delayed/subprocess failure and Stop checks passed.
- Independent clean worktree: the same scoped helper changes passed all 81
  helper tests and all 7 native Telegram tests without the user's local edits.
  The realistic capacity regression failed before this repair and passed after it.
- The packed tarball audit passed for 525 files with no warnings. That exact
  tarball was installed into an isolated prefix with lifecycle scripts disabled.
  Its actual bundled backend passed all 8 automatic/manual unary/stream checks,
  including the 218,406-token allowance against a 65,536-token remote window.
  The receipt is `primary-inference-fallback-bundle-20261012.json`.
- The verified installed-test bundle SHA-256 is
  `1c020e46538ba5a79d1d2b7586f2f8922b0f4bab36354bd878b49939461d8d9b`.

The installed 1.0.842 package remains the incident artifact until the user
publishes and updates the runtime. Publish with the usual script in the primary
checkout: `bash scripts/publish-omnius.sh`. It performs its normal patch bump.
Do not publish another package from the older source state.

No npm publication, global installation, service restart, Telegram message,
local inference or CUDA workload change was performed during this repair.
