# Telegram progress and browser-to-desktop repair

Status: complete — implemented, verified offline, and delivered to `origin/main`
Requested: 2026-10-11
Baseline: `f1b2d5f9` on `origin/main`

## Scope agreed with the user

Replace internal guidance noise and repeated request text with useful, intermittent
agent explanations in Telegram admin DMs and public groups. Reuse the main
agent's existing progress generation, preserve final-answer delivery, and reduce
the browser/desktop friction encountered while acquiring the requested images.
Include the native save-dialog boundary: expose the existing desktop controls
(xdotool backend) and configured vision/Moondream capability where authorized.

The initial assessment inspected the Copper Seal task under
`/home/roko/Documents/Projects/Adjacent/telegram_test/.omnius/`. Its first 20
retained assistant turns contained only tool calls. Read-only browser evaluations
triggered false mutation critiques. DOM summaries included JavaScript template
buttons and omitted real icon labels. A download selector matched 11 elements.
The user's subsequent observation identifies an OS file chooser/save dialog that
browser DOM inspection cannot see; this boundary must be handled explicitly.

## Work orders and delivery order

| ID | Work | Status | Depends on |
| --- | --- | --- | --- |
| [WO-01](WO-01-diagnostic-visibility.md) | Diagnostic visibility and truthful evaluation effects | Complete | None |
| [WO-02](WO-02-task-narration.md) | Shared main-agent progress for admin and group tasks | Complete | None |
| [WO-03](WO-03-browser-observation.md) | Real DOM controls, accessible names, and scoped targets | Complete | None |
| [WO-04](WO-04-download-desktop-handoff.md) | Browser downloads and desktop save-dialog handoff | Complete | WO-03 |
| [WO-05](WO-05-verification-delivery.md) | Integration verification and scoped delivery | Complete | WO-01–04 |

Each implementation commit updates its work order with behavior, checks, and any
remaining limitation. New findings enter the relevant work order; unrelated
refactors, model changes, publication, and running-session changes are outside
this repair. Do not silently expand the scope or call documentation complete
while its implementation remains open.

## Invariants

- Main-model explanations describe actual intentions, observations, and outcomes;
  no extra narrator model and no invented completion claims.
- Progress never becomes a final answer or satisfies task completion/delivery.
- Preserve chat/topic/run ownership, cancellation, steering, final review,
  attachment validation, and public/private capability boundaries.
- Keep diagnostics available without normal progress/UI noise.
- Browser observations do not assert mutations; arbitrary evaluation is not
  assumed read-only. Preserve exact target selection for interactions.
- Prefer a verified download artifact over manually driving a save dialog. When
  desktop interaction is needed, observe the actual desktop and use existing
  configured tools; do not confuse page screenshots with OS windows.
- Tests use mocked inference/transport and isolated fixtures. Live inference or
  GPU workload changes require the repository's exact-accelerator preflight.
- Preserve the original dirty worktree. Use isolated checkouts and scoped commits,
  fetch before pushing, and never rewrite shared history.

## Baseline verification

The assessment ran 163 selected offline checks: 160 passed; three Telegram
tests failed with `EISDIR` because ledger enumeration tried to read the
`presentations` directory as a file. WO-05 includes repairing those fixtures and
rerunning the affected contracts. This result is a baseline, not repair validation.

## Delivered result and limits

The integrated repair passes 445 distinct selected checks across orchestration,
Telegram progress/ownership/final delivery, browser/desktop tools, voice and
daemon transport. The recursive workspace build passes. The original three
ledger fixture failures are fixed. Expanded testing also reproduced unrelated
baseline failures in three suites; exact cases and verification commands are
recorded in [WO-05](WO-05-verification-delivery.md).

Implementation is on `origin/main`, through `346bbbea`, with scoped commits
listed in WO-05. Work was performed in isolated checkouts; the original dirty
worktree and active Telegram task were preserved. No package was published or
installed, no active service/session was restarted, and no live model, desktop,
browser or Telegram send was used for validation. Applying the new package to a
running instance and observing live model compliance are separate rollout work.
