# WO-06 — Medium-horizon browser tasks and ordinary runtime availability

Requested: 2026-10-12. Baseline: `b9a32342`. Status: source integration and
controlled evaluation verified; live-model and active-daemon acceptance open.

The user requested longer form filling, multi-page navigation, public API-document
extraction from Mapsource, and action-speed measurements. They also asked whether
the delivered controller is automatically usable in ordinary TUI and Telegram
conversations. These are separate acceptance gates.

## Baseline findings

- The visible four-action result used the production runner, tools, configured
  models and terminal verifier through an explicit host harness. It did not test
  natural-language tool selection in an ordinary conversation.
- Source TUI and admin-DM registrations expose semantic browser/native primitives,
  but create tools without the configured decision model or visual grounder.
  `runConfiguredHostUiSession` is a separate explicit CLI path. Nesting it inside
  an active conversation would bypass the intended task ownership boundary.
- The currently running daemon and global launcher use installed Omnius `1.0.837`.
  Its bundled `dist/index.js` lacks `computer-use`, `ui_snapshot`, `ui_step`,
  `native_ui`, `rapidVisualGrounder` and `omnius.ui-state.v1`. Source availability
  cannot establish availability in that running process.
- Browser snapshots previously concatenated iframe text into page text without
  document URL provenance. That aggregate cannot establish a sourced extraction.
- The historical headed harness hard-codes one plan, four decisions and one point.
  It remains a regression fixture; a new suite needs independent scenario budgets.

## Delivery checklist

| ID | Deliverable | Acceptance |
| --- | --- | --- |
| 06-A | Shared normal TUI/admin-DM browser configuration | Host-pinned policy/vision, per-request preflight, trusted task context, separate admitted `ui_step`, honest absent/unready capability, no inference on construction |
| 06-B | Longer controlled browser tasks | Rich form, stateful multi-page wizard, documentation traversal; actual owned browser and runner; independent saved values and terminal receipts |
| 06-C | Sourced public documentation extraction | Actual browser observations, exact source URL, document/snapshot hashes, supporting spans; absent, changed, truncated or iframe-only text cannot establish success |
| 06-D | Speed and request accounting | Task wall time, observe/decide/execute-and-verify phases, model request counts/receipts, startup/review separated; scripted baseline identified explicitly |
| 06-E | Bounded reusable live evaluation | Exact pins, physical preflight, one in-flight request, bounded per-case budgets, fresh approval, durable planner growth identities, owned batch cleanup |
| 06-F | Delivery and deployed-state evidence | Focused tests/build, scoped commits pushed; source/runtime/installed-daemon availability reported independently |

The normal main model remains the planner in 06-A. A separate main-model call
between steps is still required; this slice must not be described as an autonomous
small-model continuation loop. The separate host controller can run such bounded
plans and is evaluated by 06-B/06-E. Existing public/group host-input restrictions
remain authoritative. Progress delivery retains its current Telegram ownership
and routing contracts.

## Test and execution boundaries

Controlled cases use a separate CPU-rendered owned Chromium window and no user
profile. Public Mapsource work is read-only documentation traversal; it does not
require accounts, API credentials or paid endpoint requests. The source preparation
used the public [documentation](https://mapsource.io/docs),
[routing guide](https://mapsource.io/docs/navigation-api),
[quickstart](https://mapsource.io/docs/quickstart), and error/limit guides. Preparation
is not agent execution evidence.

Fixture planning plus rules selection establishes browser/executor behavior and
its latency. It does not establish model quality, natural-language discovery or
configured-model speed. Mocked model replies are integration tests only. Live
model results require new receipts and cannot reuse the previous 25.05-second
four-action measurement as a medium-horizon result.

Fresh discovery during this work reported two visible A100s and an active voice
workload, but GPU2 failed its device-handle query. The broker reported a missing
selected UUID and `admission_blocked: true` with NVIDIA recovery required. The
user correctly noted existing GPUs are online and in use; that is compatible with
new lease admission being blocked. No model load, inference or service restart
was performed on that result. Preserve existing workloads and recheck before any
live test; do not substitute the GT 1030, CPU inference or an unchecked endpoint.

## Results

The final CPU-rendered **visible** Chromium suite completed all four cases. Each
has a matching committed terminal ledger, independent evaluation, zero model
admissions/receipts, and a stopped owned browser PID. Owned-window screenshots
were captured by PID/start-time and X11 window PID identity, then inspected.

| Case | Verified actions | Execution wall | Observe / select / input-and-verify totals |
| --- | ---: | ---: | ---: |
| Rich form | 8 | 3.08 s | 1127 / 0.18 / 1321 ms |
| Multi-page wizard | 11 | 3.61 s | 1032 / 0.10 / 1975 ms |
| Local guide traversal | 5 | 1.92 s | 620 / 0.04 / 701 ms |
| Mapsource documentation | 8 | 11.26 s | 2585 / 0.11 / 7820 ms |

Each case had a separate 10-second review pause, excluded above. Selection is a
rules baseline, so its near-zero latency does not describe Kev or another model.
Startup/navigation, planning and terminal bookkeeping account for the remaining
wall time. This is one final visible sample per case, not a reliability estimate.

A preceding headless sample completed the local cases in 4.08 / 5.23 / 2.95 s.
An earlier visible public-docs sample completed in 14.46 s with a separate
30-second review pause. The public case traverses nine actual guide pages and
checks seven fixed facts; the local guide checks five. Source evidence includes
exact URLs, text/snapshot hashes and supporting spans, independently rechecked
after execution. This verifies extraction plumbing, not open-ended model reading.

Two failed attempts are retained:

- The original wizard exhausted four immediate reads before its 450 ms server
  confirmation. `RapidUiController` now observes immediately, then waits up to
  100 ms between checks with a 2500 ms verification deadline and 20-observation
  cap. Stop interrupts the wait, and input is not repeated. The repaired headless
  save verified after five observations over 466 ms.
- The public homepage changed contact-dialog nodes, chart text and its rotating
  greeting between observations. Exact revision validation rejected the first
  link before input. Ten captured states establish that drift. The documentation
  regression now starts directly at `/docs`. Broad revision relaxation was not
  introduced; recovery from continuously changing pages remains a separate gate.

The browser adapter also preserves uncertain input delivery when a click or
navigation wait throws after input may have occurred. Only known pre-input
validation failures are rejected. The controller retains the attempted input
under its idempotency/verification guard; a timed-out click cannot become proof
of absence or an automatic second click.

06-A is implemented for normal TUI, in-process subagents and admin DMs using the
shared private operator configuration. Typed discovery and current-task ownership
are integration-tested through the ordinary runner path. Construction sends no
requests. The main model still plans between actions; the autonomous small-model
continuation goal is not satisfied by this per-step integration. Public/admin
group host-input policy remains unchanged.

06-E now supports bounded per-case planner/policy/point budgets, physical request
receipts, one in-flight request, and a private fsynced planner-growth journal with
fresh request identities. Its offline admission/supervisor tests pass. The final
hardware recheck still reports `admission_blocked: true` and the same missing GPU2
UUID while the voice lease is active. **No new CUDA lease, model load, live
inference, or service restart occurred in this work.** New medium-horizon
configured-model speed/quality measurements remain open.

Verification:

- Final clean-cache all-workspace build and `scripts/build-publish.mjs` passed.
- Normal runtime/discovery/guidance/contracts: 50 CLI checks; ordinary runner
  ownership: 4; existing execution visual/cancellation/contracts: 76; Telegram
  progress/private core: 63. All passed.
- Controller timing/core: 163; harness/admission: 33; Python supervisor: 33.
  All passed. Real browser adapter regression: 84 Chromium/Firefox checks passed,
  with 42 optional BiDi cases skipped. Four affected Chromium/Firefox uncertainty
  and pre-input download validation checks then passed after the final guard.
- Actual local integration invocation passed three behaviors; its fourth
  grant-denial diagnostic assertion was corrected to inspect the retained
  outcome instead of a later recovery message, then passed in a targeted rerun.
  The final complete integration rerun passed all four tests, including actual
  CLI progress/cleanup and Stop. The four-case visible suite also passed independently.
- The broader Telegram conversation-admission suite has five failures reproduced
  against committed HEAD in an isolated loader; they occur before browser work.
  They remain baseline debt, not a green-suite claim.
- Pack audit passed for 518 files. Six shipped gate/lock assets match source
  bytes; an isolated metadata-only gate import resolves the packaged lock and
  imports no Torch. CLI help passed with the declared Zod 4 dependency. A bare
  workspace staging launch initially resolved unrelated root Zod 3 and failed;
  clean dependency installation was not tested. Nothing was published or installed.

Implementation delivery: `61aa0685` (source provenance and phase metrics),
`2c941d7b` (shipped preflight assets), `86c6881c` (bounded model suite admission),
`f4d16f47` (delayed effects and uncertain delivery), and `73687219` (normal runtime
integration), pushed to `origin/main`. The suite/evidence delivery is recorded
with this work order's final commit.

Use [MEDIUM_HORIZON.md](../../../computer-use/MEDIUM_HORIZON.md) to rerun the
suite and [RUNTIME_BROWSER.md](../../../computer-use/RUNTIME_BROWSER.md) for
operator configuration. Retained selected artifacts and their SHA-256 manifest
are in [medium-horizon-20261012](../../../computer-use/evidence/medium-horizon-20261012/README.md).
The observed installed daemon remains `1.0.837`, without these primitives; no
real Telegram delivery or natural-language live tool-selection claim is made.
