# Built-in free inference and clean Telegram settlement

Status: implemented, tested and committed. Live service limits are recorded
below rather than claimed as model availability. Publication and bot restart
remain operator-owned.

## Ordered work

1. Repair the public independent-review request. Real installed-bundle probes
   reproduced sequential HTTP 400 rejections for `uniqueItems`, then top-level
   `anyOf`. Keep the full host schema and public semantic-review gate. Permit a
   bounded repair for each newly observed supported limitation on the same route.
2. Exercise other eligible free models with real synthetic generations through
   the actual production request builders/parsers. Cover public review as well
   as simple replies. Record actual failures/quotas; never claim availability
   from mocked inference or promise that external services never fail.
3. Include the free models in the default model-selection list without requiring
   credentials or endpoint configuration. Use model names and neutral free-model
   presentation in the picker. Retain truthful transport identity in internal
   diagnostics/provenance. Preserve existing configured model choices.
4. Update first startup and lightweight setup to offer a working built-in free
   route before asking for credentials or downloading/loading local models.
   Account-based providers and local inference remain selectable options.
5. Repair the live TUI's premature/duplicate completed states and later failed
   state shown after a healthy `pong`; inspect the separate public conversation
   failure and stalled continuation path. Separate inference-phase completion
   from task settlement and preserve actual Telegram acknowledgement authority.
6. Rebuild, run meaningful regression tests and live remote acceptance, verify
   the packaged artifact, then make scoped commits and push the tracked branch.
   Publication/restart remains operator-owned.

## Current live anchors

The running bot uses 1.0.850, bundle SHA256
`d8636120a09b01e5510cb048c5cafcbb0598474653831174196e6e708de3c33e`.
Read-only Telegram history confirms private `pong` (message 3368) and public
`pong` (message 5805). Public generation succeeded; the independent
`chat-fast-path-review` request then exercised an additional schema shape that
the earlier attention/simple-reply live harness did not cover.

Private admin input can use its explicitly admitted single-pass path and avoid
that independent public review. Do not remove the public review to copy the
private behavior. The earlier schema-only `review` probe used the producer’s nominal 1,300-token
limit. The complete Telegram method applies its observable-request budget
(16,384 tokens in these remote probes), so the new all-model harness invokes
the actual attention, generation, independent-review methods and parsers.


## Implementation and regression verification, 2026-10-13

- Public and private generation share the same provider adapters. Public review
  retains its independent original host schema and semantic acceptance gate.
- Every remote JSON phase receives its full producer contract in the prompt,
  including on models with unknown or supported native structured output.
  Explicitly unsupported native formats use prompt-only generation with original
  host validation, complete-response buffering and unchanged tool authority.
- Big Pickle’s exact Chat identity has an observed native-format override:
  native requests returned HTTP 400 `{"model":"big-pickle"}`; omitting only that
  format gave HTTP 200 generation and review with host acceptance. Production
  revalidation then passed attention, generation and review without the probe flag.
- Cat Jam’s bounded diagnostics learn `uniqueItems` and a root conditional-union
  restriction. Only the native grammar is lowered. Nested closed objects, identity,
  authorized tools and coupled fulfillment/continuation checks remain unchanged.
- Exact pinned-gateway HTTP 429 `Endpoint is unavailable` errors retain their
  HTTP status but are classified as upstream outages, not quota exhaustion.
  They cool that model without scheduling a quota retry. Genuine rate limits keep
  their existing cooldown and retry policy.
- Remote streaming keeps the backend’s 45-second raw-byte idle watchdog and
  the 120-second aggregate absolute ceiling. Hidden reasoning and buffered JSON
  no longer hit a second 45-second absolute timer. Shorter caller limits and Stop
  still apply; unary requests retain their original absolute attempt bound.
- Built-in selected models enter the same catalog, transport, schema adapters,
  validation and lifecycle reporting used by fallback. Explicit selection is first;
  automatic alternatives use the capability ordering. Manual mode asks before
  alternatives; disabled fallback still permits the selected built-in model.
- New installations use built-in free inference with no key or local download.
  `/model`, Telegram’s isolated model picker and completion include the neutral
  `Built-in free` category. Bundled choices remain visible offline with truthful
  availability labels; live pricing/retirement still control admission.
  Existing local/keyed profiles and explicit transport overrides remain intact.
- Independent regression checks: 504 orchestrator, 254 CLI and 15 provider registry
  tests passed. These use controlled network fixtures for invariants and are
  separate from the live model evidence. Full workspace rebuild passed.
- The compiler-free fresh npm install passed for the packaged candidate, including
  SQLite and AIWG checks; artifact audit passed and the 566-entry tarball has no
  sourcemaps. No version was published and the running bot was not restarted.

The existing README generated REST inventory check fails at the unchanged REST
block (the task’s README diff only updates setup/prerequisites). The independent
REST source check passes; this pre-existing inventory drift is not included in
this repair.

Live coverage deliberately sends only invented messages through real Telegram
methods and providers. It never calls the Telegram Bot API, dispatches a native
host tool, reads private conversation history, or invokes local inference. Thus
these receipts establish inference-path behavior, not a new delivery ACK or
native cross-application action test. Delivery/settlement regressions are in
`87cecd19`; the installed 1.0.850 bot remains the operator’s running version.


## Live packaged results and delivery

Implementation: `b94a6179` public review, `87cecd19` Telegram settlement,
`273a7bdb` built-in free integration and common adapters, `65a15404` Step’s
opaque complex-schema compatibility. Commits are delivered to `origin/main`.

The [durable live receipt](evidence/telegram-all-free-live-20261013.json) records
all nine eligible models, exact bundle/catalog hashes, phase outcomes, safe
provider diagnostics, original host acceptance, and physical request counts.
No inference reply or catalog response in these live runs was mocked.

| Model | Live result through Telegram methods | Adapter or remaining limit |
| --- | --- | --- |
| Big Pickle | Attention, generation and independent review passed in installed sweep | Exact Chat native-format opt-out; original host contract retained |
| Cat Jam | All three phases passed in installed sweep | Learned `uniqueItems` and root conditional-union restrictions |
| Ling 3.1 Flash | Earlier attention/generation passed; final isolated probes returned endpoint-unavailable HTTP 429 | Declared native-format opt-out; complete installed Ling→Muse cascade passed |
| LongCat 2.5 Preview | All three phases passed in installed sweep | Full original JSON guidance, native Chat transport |
| MiMo V2.6 Flash | All three phases passed in installed sweep | Full original JSON guidance, native Chat transport |
| Muse Spark 1.3 | All three phases passed in installed sweep | Responses auto-only tool choice and learned native schema restrictions |
| Nemotron 3.5 Lightning | All three phases passed in installed sweep; attention took 61.9 seconds under default deadlines | Raw-byte idle watchdog plus aggregate absolute bound; no invented thinking flag |
| Space Bunny | All three phases passed in installed sweep | Full original JSON guidance fixed the prior JSON-syntax failure |
| Step 5 Preview | Initial attention passed; patched final installed generation and independent review both passed (HTTP 200/200) | Simple native schema works, complex schema returned opaque HTTP 400; exact Chat prompt-only policy preserves full host validation |

Step’s final attention probe and Ling’s final probes also recorded intermittent
endpoint-unavailable HTTP 429s. These are external availability limits. The final
installed Step→Muse public cascade completed with HTTP 429→200→400→400→200:
Step’s outage, Muse generation, then two bounded review-schema repairs and an
accepted independent review. No unavailable-model quota retry or fabricated
answer was used. Both selected-model isolation and real automatic cascade modes
are available in `scripts/live-telegram-free-models-harness.mjs`.

The final unpublished candidate bundle SHA256 is
`93c9f8bae03b447f603c3245a796b32b7ee51da12389a99929489dcb2335ca27`.
Its compiler-free fresh install, SQLite/AIWG checks, 566-entry tarball audit,
README wiring and sourcemap exclusion passed. The staging version remains
1.0.850; the operator’s publish workflow must choose the next release version.
The running production bot was not replaced by these harness installs.
