# Public research stream recovery and meaningful progress

Status: repaired and verified in the isolated installed runtime. Publication
and the operator's bot restart are separate.

## Live incident

Installed Omnius 1.0.855, main PID 1141122, started at 20:10:15Z from
`telegram_test`. OMNIUS group `chat:-1003888256865`, input 5826 asks for this
week's biggest AI research; task `tgtask-b3b2802ea8259be5a181`, bridge run
`telegram:8928c8a1534bdb23`, actor
`telegram-c34a3d7d53728329-1791922362455-1`.

- Attention succeeds on Muse at 20:12:41Z after guarded schema adaptation.
- Three searches succeed. Four page fetches run: three succeed, one gets 429.
  The final successful fetch ends at 20:13:03Z.
- The next actor request has approximately 34,904 input tokens. Its context
  falsely says `web_fetch:no-target` has failed 46 rounds, despite this being
  a fresh native goal with one failed fetch.
- At 20:15:03Z finalization commits `inference_stream_interrupted`, turn 3,
  `retryable:false`, before any accepted final answer. This is approximately
  120 seconds after the last request started. The old actor trace was not saved,
  and the interruption error discarded its cause/frame categories, so the
  precise provider failure or timer cannot be reconstructed from this receipt.
- Message 5828 is solely the canned incomplete sentence. An absent native goal
  assessment caused the host to append it to an empty failed result, then mark
  intake failed after delivery. A separate empty-result branch would instead
  wrongly consume it as a social no-reply if the appendix alone were removed.
- The first edited progress message stores three differently worded “pulling
  the research” paragraphs. Exact-text deduplication permits near duplicates;
  no stage grouping replaces the stale explanation when inference waits/fails.

Incident bundle SHA256:
`1c51f6f647513d3b3f43c2ed42d9aa3ea3bb56eeb0a27aab4daf63cdd85e5625`.
Finalization SHA256:
`362776f300db24c47c81b32708ac11ea90151aedb1e7996c37029904e6d4e9dd`.
Scoped log copies and hashes were preserved outside the live state. No
historical transcript or private source content is sent by the live harness.

## Repairs and constraints

1. Scope persisted convergence to stable native task identity and epoch;
   retain genuine continuations, isolate new original inputs. Hash the whole
   normalized URL as the operation target, preserving query distinctions and
   omitting fragments. Successful work resolves only its exact target/family.
2. Buffer each Telegram primary generation until its complete turn validates.
   Failed uncommitted text/tool frames may fall back without replaying prior
   successful research tools. Use the existing configured per-attempt limit
   within the total/caller limit, preserving time for the next route. Immediate
   TUI streams retain their no-replay fence after actual output. Stop and manual
   approval remain authoritative; marked auxiliary review allowances remain.
3. Hidden reasoning/accounting alone never constitutes an irreversible answer
   or tool effect. Discard failed hidden-only attempts before switching models.
   Bound buffered frames/bytes, and recheck ownership before release.
4. Save actor inference routes as well as auxiliary routes. Record only static
   failure category, status, deadline kind, frame categories and commit policy;
   never provider bodies, generated text, endpoints or credentials.
5. Group public progress by actual host tool stage. Replace same-step model
   explainers, retain factual completed stages and show waiting/paused/stopped
   state. Counts describe successful operations, not unique sources or goal
   fulfillment. Keep tool chronology behind Activity and explicit progress
   available; do not add progress-only model requests or change admin behavior.
6. Empty failed native work produces no fabricated final answer, no final
   delivery receipt and no consumed intake. Retain the original goal/evidence
   and make the existing progress message paused. Legitimate current accepted
   partial answers retain their separate receipt-backed delivery path.
7. Hand the bridge an issued, immutable source receipt directly from the
   actual runner result, bound to the current scope/run/epoch/tool action.
   Display projection cannot remove verified source custody. Forged, copied,
   stale, failed or blocked event carriers confer no source evidence; legacy
   marker parsing remains only for events without the new field.

The follow-up [complete-source work order](WO-web-source-full-coverage-extraction-20261013.md)
tracks the operator's additional requirement to remove page prefixes and
account for every source chunk when dedicated extraction is necessary.

## Verification

The isolated workspace build, tarball audit (573 files, no warnings) and fresh
compiler-free npm install pass. Isolated suites pass 137 orchestrator checks,
133 CLI checks and 40 typed receipt/public-native checks (overlapping suites,
not a unique total). Four broader native-transport fixture failures reproduce
unchanged on baseline `ba291791`; they are not represented as passing.

The [installed live receipt](evidence/telegram-public-research-stream-live-20261013.json)
records nine real free-model POSTs, two actual scoped RFC page reads and real
native goal assessment. After both reads, the harness cuts a real Step text
delta (227 bytes forwarded). MiMo completes the same logical generation,
without another source read. Both issued receipts reach the bridge, both source
URLs reach the final wire text, and intake settles consumed/goal fulfilled.
Bundle SHA256: `666eaaf77fce81bd82f2a214c0f787447af75decaafc6bd2bda65527822a7266`.

Only Telegram transport/ACK is simulated in disposable state. No inference or
catalog response is fabricated, no local inference runs, and no actual Telegram
message is sent. Earlier fault-fixture runs completed useful research but did
not actually inject a disconnect; they are excluded from interruption proof.
The first fresh-install attempt hit a SQLite prebuild download timeout; the
unchanged tarball passed on retry. Actual Telegram delivery/typing needs
operator deployment. The separate full-source extraction work remains tracked
in its follow-up order.
