# Recover accepted Telegram drafts without restarting completed work

Status: implementation and installed-package live verification complete.
Follow-up to the Case Theory stall on installed Omnius 1.0.854, PID 161677,
running from `telegram_test`. Publication and bot restart remain operator-owned.

## Incident anchors

The original input is `telegram:chat:-1003960149435:0:1083`; task
`tgtask-2d2fee541b8304db4568`, bridge run `telegram:2a5a22214cf84400`,
actor run `telegram-84e4e6e5411c9753-1791918893130-1`.
The input explicitly addresses the bot while replying to an independently
adopted message 1081 by another participant. Private source and draft text are
not copied here or sent to the synthetic live test.

- 19:15:01Z: Telegram progress delivery succeeds.
- 19:15:39Z: `task_complete` accepts a full answer for deferred Telegram
  delivery. The actor did draft an answer; this was not a generation failure.
- Finalization `terminal-3e4feb48f9f23fd29a87` records
  `incomplete_verification`: required observation and visible response have no
  canonical receipts. The source-answer review's support verdict is separate
  from actual observation or provider delivery evidence.
- The old runner exits immediately on the accepted completion despite missing
  work, filters the full answer from its return, and gives terminal review an
  empty candidate. Review requests another action run.
- Continuation reopens the unchanged source and falsely rejects it as
  `original_native_source_changed`: it compares every adopted source's author
  to the current task owner. Initial admission had correctly preserved the
  independently adopted participant's author. Both native envelopes and text
  hashes still match their immutable spool/requirement bases.
- At 19:22:02Z the exhausted planning transition becomes
  `reconciliation_required`. `missing_action_tool_names` is the generic
  diagnostic after the host removes a rejected plan, not evidence that the
  model independently omitted tools. Public chat gets no paused status.
- A later human follow-up 1086 obtains assistant reply 1089 at 19:30:36Z. The
  original request required another prompt to recover; it was not deliberately
  ignored by the attention router.

The installed incident bundle SHA256 is
`1c385603d694ff5ca04edf20af158b1f75895e96b900e603a9063112fe2a257c`.

## Ordered repairs

1. Validate retained adopted sources against each source's admitted actor,
   envelope, text basis, topic and immutable custody. The original task owner
   still owns the task. Legacy actorless locators retain the owner-only rule.
2. Before leaving an accepted completion, check required non-provider work
   effects. Keep the same runner, tools, workspace and exact answer while
   gathering only the missing evidence. Use the existing bounded completion
   budget. Do not hold for visibility or provider delivery alone.
3. On bounded exhaustion, retain the exact accepted draft as typed
   `pending_work_evidence` metadata. It cannot become a deliverable, successful
   observation, provider ACK or completed objective. Reopen the exact committed
   terminal marker and native source custody before exposing it to review.
4. For an already admitted public human task whose planning recovery stops,
   send one brief paused status with the original issued reply grant/topic.
   Preserve Stop, later-human preemption, native ownership, durable attempt
   deduplication and physical ACK checks. Keep status separate from task result
   and effect receipts; never broadcast ambient or peer-bot failures.
5. Verify isolated source, full package rebuild, tarball and fresh install.
   Run the installed production runner with real remote free-model responses
   and a real source read on synthetic data. Make a scoped commit and push.

## Source locations

- `packages/orchestrator/src/agenticRunner.ts`: completion reply ownership,
  required-work hold, bounded retention and result handoff.
- `packages/cli/src/tui/telegram-bridge.ts`: retained native source validation,
  terminal review draft context and public paused-task ownership/dispatch.
- `packages/cli/src/tui/telegram-planning-failure-notice.ts`: fixed public status
  and existing durable exclusive attempt store.
- `scripts/live-telegram-retained-work-harness.mjs`: synthetic installed-runtime
  live acceptance. It pins the eligible Muse endpoint, bounds requests, forbids
  local inference and Telegram sends, and invents no model/catalog replies.

## Verification

All 66 new regressions passed: 13 runner repair, 11 adopted-source custody,
9 retained-draft Bridge custody, and 33 public paused-status cases. The isolated
orchestrator batch passed 62/62. The broader isolated CLI batch passed 199/201;
its two old source/owner guard final-state assertions fail identically on
untouched, fully compiled `ba291791`. An additional existing terminal-review
fixture fails on that same baseline (17 other cases pass). These are recorded
test limitations, not an assertion that the entire suite is green; no custody
guard or test assertion was relaxed to obtain success.

The full isolated workspace rebuild, 571-file tarball audit (zero warnings),
and compiler-free fresh npm installation passed. The installed candidate uses
the tracked test manifest version 1.0.790; it is identified by bundle SHA256
`c78a20f7d05ef313e2226211235fb57402a52249aa366b7274a8a1cc28479a5f`,
not an npm publication or replacement of the operator's runtime.

The real installed Muse run passed in 16.7 seconds, with three HTTP 200 streamed
Responses requests and the sequence `task_complete → file_read → task_complete`.
Exactly one bounded required-work hold preceded one actual production
`FileReadTool` read. Both completion submissions have the same exact answer
hash. Observation becomes satisfied; only visible response remains pending.
The runner returns `provider_delivery_pending`, no retained-only draft, no
accepted/completed result, and zero provider ACKs. Model/catalog replies were
not mocked; all source content was synthetic and no Telegram message was sent.

This live test covers the installed runner and real source tool with a
host-owned scratch acceptance adapter. It does not exercise real Telegram
ingress/reply grants, the Bridge source-answer critic, or Telegram API delivery.
Actual paused dispatch/source custody are covered by the focused synthetic
transport tests. Actual Telegram delivery and typing visibility require the
operator's updated publication/restart; no live conversation state is modified
by this repair or test. [Safe receipts](evidence/telegram-retained-source-handoff-live-20261013.json)
retain hashes and synthetic diagnostics without private source/draft text.

Architecture: [separate review, delivery and task states](../../.aiwg/architecture/adr-completion-review-delivery-states-20261013.md).
