# Live native-planning cascade and TUI responsiveness

Status: implementation and live packaged acceptance complete. The operator
reported publication in progress; runtime restart/activation is not yet
verified. Follow-up to the live 1.0.851 report after the built-in free-model
integration; do not substitute another ping-only check.

## Actual live anchors

- PID 338848, working directory `telegram_test`, version 1.0.851, bundle SHA256
  `93c9f8bae03b447f603c3245a796b32b7ee51da12389a99929489dcb2335ca27`.
  This matches the previously validated installed candidate: stale deployment
  does not explain the discrepancy.
- Public input 1002 in the operator’s Case Theory chat entered processing at
  2026-10-13T08:38:34.455Z. The next input 1003 entered processing at
  08:44:23.071Z and produced a valid observer/no-reply decision at 08:44:39.619Z.
- Read-only inspection of the live terminal’s expanded inference block found
  `router-native-requirements`, 272 retained control rows, 4m37s duration:
  local HTTP503 → Muse HTTP400 twice → Step upstream-unavailable HTTP429 →
  Ling upstream-unavailable HTTP429 → MiMo, with first output reported at 277.2s.
  The display reported successful completion with 1207 tokens. Provider HTTP400 bodies
  were not retained in the existing durable logs.
- The separate 16s router block supplied by the operator is a successful
  Muse schema repair, followed by a valid no-reply decision. It is not the
  failing native cascade.
- At inspection, TUI stdout/stderr are only `/dev/pts/6`; current provider-phase
  diagnostics are not persisted. Existing root crash/session logs predate
  this restart. Broker access logs show 422/503 local chat failures without
  model/request correlation, insufficient to classify those refusals.
- Process sampled at ~1GB RSS and ~12 minutes CPU in 20 minutes of uptime; current
  updates and accepted replies show it is not permanently deadlocked.
  A 3.3GB deferred-memory spool exists, but inspected append/maintenance paths
  do not read it wholesale; do not attribute the freeze to it without proof.

## Ordered repairs and acceptance

1. Reproduce the actual native-requirement producer and independent critic
   through the installed bundle using invented source text and active tasks.
   Keep the full original host schema/owner/scope validator and prohibit
   local inference, native action dispatch and Telegram API sends. Capture
   bounded safe upstream failure diagnostics for these synthetic requests.
2. Repair evidenced native grammar/request incompatibilities at the shared
   phase boundary. Preserve host semantic validation, capability/authority,
   provenance and original contract coupling. Validate real responses; no
   mocked availability claim and no unconditional HTTP400/422 bypass.
3. Remove repeated formatting of unchanged retained inference transcripts
   during TUI repaint. Preserve full read mode, resize, thinking visibility,
   title time/blink, retention bounds, phase reuse and input handling. Test
   repeated repaint work with large completed native transcripts.
4. Add safe durable phase/route diagnostics so future live failures can be
   traced without extracting terminal rows. Do not persist credentials,
   endpoint keys, raw user prompts or private generated transcripts. Optional
   display observers must not fail/replay an inference request.
5. Run relevant regressions, real native proposal/critic acceptance, rebuild
   and make scoped verified commits to the tracked branch. Preserve all
   unrelated dirty changes. Publication/restart remains operator-owned.

The previous all-model ping harness established attention/reply/review behavior
at healthy moments; it did not cover native task projection/planning. This
follow-up closes that coverage gap rather than reasserting the ping results.

## Reproduced causes and repairs

The installed native producer reproduced `uniqueItems` rejection, then root
`anyOf` rejection twice. The optional-field repair exposed the next exact
upstream error: every property must be required (`activeScopeId` was named).
The host contract deliberately permits omitted fields; making them required
would alter its semantics. Recovery now learns that precise incompatibility
per model/endpoint/API, omits only native schema framing, and preserves the
original full host schema, parser, tool arguments and independent critic.
An adaptation must change the previous wire schema before it can justify
another retry. Recovery remains bounded by three distinct restrictions.

A controlled prompt-only request also exposed true raw-stream inactivity:
HTTP200 followed by the backend's 45-second idle watchdog. It was incorrectly
reported as generic transport failure. Native planning/review streams now
allow 90 seconds under the default policy, capped by caller/aggregate deadlines;
explicit timeout settings and ordinary chat/unary/local paths retain their
limits. Exact watchdog failures report `timeout`. No undocumented reasoning
flags were added.

The controlled 90-second diagnostic accepted a real 7-obligation/two-method
proposal through the original parser and admitted an independent critic.
This diagnostic intentionally omitted native framing at the wire; it is
recorded separately from automatic production-adapter acceptance.

[Safe live receipts](evidence/telegram-native-planning-live-20261013.json)
contain hashes/metadata for invented fixtures, not private chat transcripts.

## Delivered coherent repairs

- `e65ed677`: cache unchanged TUI inference transcript rows; 61 focused checks
  and CLI build passed. This removes repeated formatting work; it does not
  establish that repaint was the sole cause of the reported freeze.
- `04b4c5e7`: private bounded route traces at
  `.omnius/telegram-inference/routes.jsonl` and isolated optional observers.
  Files use 0600, the directory 0700, pending observations 128 and files 5 MiB
  with one rotated archive. Trace failure cannot fail/replay the request.
  Backend observer 14, trace 20, Bridge integration 22 checks passed; the
  combined focused CLI run passed 105 checks. Stop/custody fences remain.
- `42419cd2`: precise optional-schema recovery, ineffective-retry prevention,
  scoped default native stream allowance and correct timeout classification.
  Automatic compiled-workspace live acceptance passed both native phases:
  three distinct HTTP400 capability errors repaired on Muse, then original
  proposal validation and independent critic acceptance, without model failover.
  Final focused verification passed 531 orchestrator + 112 CLI = 643 checks.
  Full workspace rebuild, artifact audit and compiler-free fresh npm install
  passed. At verification, PID 338848 still uses the old `93c9f8...`
  bundle; this pass has neither published nor restarted it.

## Broader regression limits

The expanded run found four transport-fixture and two native-intake guard
failures. An isolated source archive of `e65ed677`, with the original Runner
rather than dirty/compiled workspace output, reproduced all six failures.
Three transport fixtures omit the previously required
`observationSourceMessageIds`; one ambiguity expectation predates the
advisory-policy change. The two intake guard cases have an existing objective
contract retention discrepancy. They are recorded as baseline failures and
were not silently changed to claim a green full suite.

## Final installed-package acceptance

The freshly installed tarball used bundle SHA256
`36c6a7172915a3edc9dc2a355ee9628f64a2d865e8e0cc3547b10e88d37a45e2`.
A fresh process automatically recovered three distinct schema HTTP400s on Muse
and retained the selected model throughout. The native proposal took 69.1 seconds,
parsed five obligations/two methods with new-task continuity; the independent
critic took 103.9 seconds overall, including 64.6 seconds before HTTP headers,
and was admitted with method verdict `accept`. Both phases retained the
original host schemas. Five actual generation POSTs were made; no model or
metadata reply was mocked, and no local inference, native tools or Telegram
API sends were performed. This validates production inference/validation
boundaries, not a fresh Telegram delivery ACK or cross-application task.

Reproduce against a fresh isolated installation:

```sh
node scripts/live-telegram-native-requirements-harness.mjs --run \
  --bundle /path/to/isolated/install/lib/node_modules/omnius/dist/index.js \
  --output /tmp/omnius-native-planning-live.json
```

The default harness pins Muse and denies alternative models, local inference,
Telegram API calls and native action dispatch. `--inspect-only` makes no
network requests. Optional diagnostic flags are recorded separately and are
not substituted for the automatic installed-package acceptance above.

The operator reported npm publication in progress after installed-package
acceptance. A second rebuild, intended to refresh packaged documentation,
was stopped in workspace compilation before it could replace `publish/`.
The earlier complete build/audit/install/live validation remains the evidence
for the code bundle above. No further publish-directory mutations were made
by this pass while the operator published.
