# WO-05 — Verification and delivery ledger

Status: in progress.

Original worktree: `/tmp/omnius-rapid-computer-use-20261011`, starting at `4ba37c16`.
After a host restart cleared `/tmp`, persistent continuation uses
`/home/roko/Documents/Projects/Adjacent/omnius/omnius-rapid-computer-use-20261012`
from delivered `291f8373`. Uncommitted draft tests from before the restart are
historical only; the replacement implementation is tested independently.
Preserve the original dirty workspace and active Telegram task. The earlier
repair passed 445 selected checks plus recursive build; its separately reproduced
baseline suite failures are recorded in its WO-05, not silently attributed here.

Run each meaningful regression, then package/workspace build and the affected
Telegram/browser/cancellation/receipt contracts. Any frontend transport change
also requires the repository's daemon lifecycle gate. Unit and local-browser
fixtures must not contact local inference or the user's desktop.

Record exact commands, case counts, failures, commits and limitations here after
execution. Commit only task-owned paths after each coherent verified repair;
fetch and push to tracked `origin/main` without rewriting history. Source changes
are not installed into the active task merely by pushing them. Model rollout,
native platform validation and architecture phase closure need their own evidence.

## Runs

- Dependencies installed offline with the frozen lockfile and lifecycle scripts
  disabled. This is setup, not implementation verification.
- `pnpm --filter @omnius/orchestrator... build`: passed (7 dependency projects).
- Recovery: `pnpm --filter @omnius/orchestrator exec vitest run
  tests/observationLedger.test.ts tests/negative-observation-recovery.test.ts
  tests/agenticRunner-negative-observation-recovery.test.ts
  tests/agenticRunner-control-observation.test.ts tests/focusSupervisor.test.ts
  tests/discovery-advisory-progress.test.ts
  tests/action-frontier-recovery.test.ts`: 117 passed.
- `pnpm --filter @omnius/execution exec vitest run
  tests/shell-receipt.test.ts`: 10 passed.
- Transform/session-lock modules: 20 passed. Decision adapters and existing
  desktop-click suite: 53 passed. These components are still being integrated.
- At this initial checkpoint, no live model or desktop tests had run. Subsequent
  isolated platform and model runs are recorded below; WO-04 remains mandatory.

Native-save integrated check:
`pnpm --filter @omnius/execution exec vitest run tests/desktop-click.test.ts
tests/vision-action-loop.test.ts tests/native-save-verification.test.ts
tests/rapid-ui-geometry.test.ts tests/rapid-ui-desktop-session.test.ts`:
**90 passed** (31 + 22 + 17 + 15 + 5). Execution typecheck passes.

Persistent controller verification (2026-10-12):
`pnpm --filter @omnius/execution exec vitest run tests/rapid-ui-core.test.ts
tests/rapid-ui-visual.test.ts`: **63 passed** (47 + 16), independently rerun.
`pnpm -r build`: all 11 workspace projects passed. Visual tests use host callback
fixtures, perform no inference/input, and do not establish Moondream accuracy.
They cover unchanged/changed captures, nonzero origins, disabled/hidden/sensitive
semantic controls, strict model identity and cancellation during capture/grounding.

Persistent adapter checks:

- Browser fixtures: `OMNIUS_TEST_PLAYWRIGHT_MODULE=/home/roko/Documents/Projects/Adjacent/noclip-unified/earth/node_modules/playwright-core
  pnpm --filter @omnius/execution test -- tests/rapid-ui-browser.test.ts
  --testTimeout=15000 --maxWorkers=2 --minWorkers=1`: **82 passed**. A final
  scroll/navigation subset passed 4/4 after adding scroll state to revisions.
  Driver 1.61.0, Chromium cache 1228 and Firefox cache 1532; fresh owned contexts.
- Existing execution tests `playwright-browser-{contract,cancellation,observation,download,install}.test.ts`:
  **39 passed**. No inference, installed browser downloads or user-browser access.
- Native suites `rapid-ui-native.test.ts`, `rapid-ui-native-gtk.test.ts`,
  `rapid-ui-native-uia.test.ts`: **35 passed**. The PowerShell test validates the
  real parser and C# interop compilation only; it invokes no Windows desktop API.
- CLI suites `telegram-desktop-tool-discovery`, `tool-contract-preservation`,
  `telegram-admin-progress`, `telegram-progress-message`, `native-inprocess-agent`,
  `telegram-private-native-runner`: **112 passed**. The final 24 initially failed
  setup because offline install skipped native build scripts; an existing exact
  better-sqlite3 12.10.0 ABI137 binding was validated/reused only in this worktree,
  then all 24 passed. No product workaround was added.
- Orchestrator suites `tool-search-desktop`, `discovery-advisory-progress`,
  `tool-batching`, `adversaryStream`, `activity-diagnostic-visibility`,
  `tool-search-schema-budget`, `telegram-progress-accounting`: **105 passed**.

Controller and runner integration follow-up:

- Added stalled decision transport/body, malformed effects, receipt custody,
  missing native application identity, pause/cancel, verifier failure and visual
  dispatch denial regressions. **164 passed** (132 core + 32 visual).
- The scoped V8 run covers six modules (contracts, controller, decision, planner,
  workflow, visual): **99.77% statements/lines, 100% functions, 95.42% branches**.
  This is not whole-package or live-model coverage. Exact command:
  `pnpm --filter @omnius/execution exec vitest run tests/rapid-ui-core.test.ts
  tests/rapid-ui-visual.test.ts --coverage.enabled
  --coverage.include='src/rapid-computer-use/{contracts,controller,decision,planner,workflow,visual}.ts'
  --coverage.reporter=text --coverage.reporter=json-summary --coverage.reporter=json
  --coverage.reportsDirectory=../../computer-use/evidence/core-coverage-20261012`.
  Reports use repository-relative paths without changing measured totals.
- The real CLI adapter lost the receiver of stateful action-intent hooks; binding
  the hook preserves its controller and exact candidate identity. Native/browser
  tools also return action effect receipts so verified inputs settle as applied
  while uncertain dispatch remains unknown. The real runner regression in
  `tool-contract-preservation.test.ts` passes **17/17**, after recursive rebuild.
  Its backend throws if inference is attempted; the test performs no live input.
- Further receipt review reproduced a concurrent duplicate returning `absent`
  while the first request was awaiting post-dispatch verification. Recorded
  outcomes now take precedence, and other in-flight requests remain unresolved.
  Eleven dedicated evidence regressions pass; combined core/visual/evidence is
  **175/175**, with execution typecheck passing. Refreshed six-module coverage is
  **99.77% statements/lines, 100% functions, 95.45% branches**. The refreshed
  coverage command adds `tests/rapid-ui-evidence.test.ts` to the command above.

Firefox BiDi follow-up:

- The fresh workspace build passes. `rapid-ui-browser-bidi` plus the five
  existing browser regression files pass **64/64** (25 transport/config/guard).
- With `OMNIUS_TEST_FIREFOX_BIDI=1` and the installed Playwright module above,
  the changed metadata, custody, authority and stop fixtures pass **21/21**
  across Chromium, Firefox Juggler and Firefox BiDi. The earlier full BiDi run
  passed 40 workflows and exposed a Blob-download timeout. Its replacement
  regression verifies explicit rejection before input for that unsupported path.
- Rebuilt CLI `tool-contract-preservation`, `telegram-desktop-tool-discovery`,
  `native-inprocess-agent`: **36/36**. Orchestrator `tool-search-desktop`,
  `tool-search-schema-budget`, `adversaryStream`,
  `activity-diagnostic-visibility`: **47/47**.
- The [owned protocol probe](../../../computer-use/evidence/bidi-blob-probe-20261012.json)
  records Firefox 151 emitting completed Blob events with no navigation ID while
  Playwright 1.61 surfaced zero download events. This is a documented capability
  limit, not a successful file-custody result.

Persistence and trace follow-up:

- `pnpm -r build`: all 11 workspace projects pass.
- `pnpm --filter @omnius/execution exec vitest run tests/rapid-ui-core.test.ts
  tests/rapid-ui-visual.test.ts tests/rapid-ui-evidence.test.ts
  tests/rapid-ui-persistence.test.ts tests/rapid-ui-trace.test.ts
  --coverage.enabled
  --coverage.include='src/rapid-computer-use/{artifact-store,workflow-store,trace}.ts'
  --coverage.reporter=text --coverage.reporter=json-summary --coverage.reporter=json
  --coverage.reportsDirectory=../../computer-use/evidence/persistence-coverage-20261012`:
  **232/232 pass**. The three new modules report **100% statements/lines/functions,
  84.88% branches**. Reports are normalized to repository-relative paths.
- Python harness: `python3 -m unittest discover -s scripts/rapid-computer-use
  -p 'test_model_gate.py' -q`: **49 pass**; `test_point_benchmark.py`: **20 pass**.
  Frozen inputs comprise ten finite-action cases and twenty generated captures.
- The isolated A100 Qwen baseline produced ten valid choices and six label matches.
  This is a synthetic finite-choice smoke result, with no GUI input or full task
  completion. Exact placement, runtime, corpus, latency, raw responses and clean
  lease release are recorded in WO-04 and `computer-use/evidence/`.

Laya staged-service follow-up:

- `.omnius/rapid-model-eval/laya-venv/bin/python -m unittest discover
  -s scripts/rapid-computer-use -p test_laya_service.py -v`: **35/35 pass**;
  Python compilation passes. No CUDA or inference occurs in these unit checks.
- The independently gated live finite-choice run returns **10/10 valid**, **4/10
  exact labels**, with native probabilities and the unfavorable errors retained.
  This result does not promote the model or claim an executed UI workflow.
- Owned-service shutdown, freed CUDA allocation and broker lease release were
  independently confirmed. Exact runtime and cleanup evidence are in WO-04.

Kev and Photon preparation follow-up:

- Kev staged service: **33/33 offline tests** and Python compilation pass. Its
  independently gated ten-case live run returned ten valid responses and five
  label matches; unsuitable choices remain visible in the evidence.
- Process-scoped VRAM sampler: **13/13 offline tests** and compilation pass.
  Kev produced 548 valid observations; cleanup verified its lease/allocation gone.
- Photon staged service: **42/42 offline tests** plus 69 shared gate/capture
  checks pass. Its exact 10.5 GB checkpoint and tokenizer hashes, 64 dependency
  versions and 516 runtime files are recorded. This preparation alone makes no
  claim of successful CUDA startup or point accuracy.
- Photon compatibility repair: **50/50 offline service tests** plus **69/69
  shared harness tests** pass. The separately gated live retry completed one
  warmup and **20/20 synthetic target hits**, with exact library mappings and
  current runner placement verified. All owned CUDA allocations and leases were
  released. The earlier startup failure remains in the evidence.

Bounded host-session integration:

- `pnpm -r build`: all eleven workspace packages pass.
- `OMNIUS_TEST_PLAYWRIGHT_MODULE=/home/roko/Documents/Projects/Adjacent/noclip-unified/earth/node_modules/playwright-core
  pnpm --filter omnius exec vitest run tests/computer-use-cli-browser.test.ts
  tests/computer-use-host-session.test.ts tests/computer-use-command.test.ts
  tests/index.test.ts tests/tool-adapter-delivery-receipts.test.ts`: **97/97**.
- `pnpm --filter @omnius/orchestrator exec vitest run
  tests/host-tool-session.test.ts tests/tool-schema-execution.test.ts
  tests/interruption-lifecycle.test.ts tests/objective-novel-operation-admission.test.ts
  tests/task-objective-operation-identity.test.ts
  tests/agenticRunner-objective-control.test.ts`: **105/105**.
- `pnpm --filter @omnius/execution exec vitest run tests/rapid-ui-core.test.ts
  tests/rapid-ui-visual.test.ts tests/rapid-ui-evidence.test.ts
  tests/rapid-ui-persistence.test.ts tests/rapid-ui-trace.test.ts
  tests/playwright-browser-contract.test.ts tests/playwright-browser-cancellation.test.ts
  tests/playwright-browser-observation.test.ts tests/playwright-browser-download.test.ts
  tests/playwright-browser-install.test.ts`: **275/275**.
- Total: **477 passing checks**. Two built-CLI headless browser cases verify the
  shipped two-action example's completed durable ledger and terminal receipt,
  and SIGINT cancellation with browser shutdown. Other regressions cover held
  validation crossing ownership, unresolved effects, rejected/malformed terminal
  verifiers, storage failure, exact planner constraints, unique model request IDs
  and denial of model-proposed action/effect widening. No model inference or
  user-desktop access occurs in these checks. `git diff --check` passes.

Isolated live host-session harness preparation:

- `node --test scripts/rapid-computer-use/host_session_smoke.test.mjs`: **17/17**.
- From `scripts/rapid-computer-use`, `python3 -m unittest test_ollama_growth.py`:
  **13/13**. Node syntax, Python compilation and whitespace checks pass.
- The fixed hash-pinned host bridge checks current placement/approval, prepares
  the owned GPU lease before one planner request, restores readiness only after
  successful completion, and stops the owned server on uncertainty. The launcher
  bounds TERM/KILL cleanup of its own worker/browser process group and records
  failures explicitly. The 30 offline checks include an owned CPU process that
  ignores TERM; they perform no model load, inference or user-desktop input.
- Exact source hashes and checks are retained in
  `computer-use/evidence/host-smoke-offline-20261012.json`.

Planner response-contract repair after the first live rejection:

- `pnpm -r build`: all eleven workspace packages pass.
- The same five-file CLI command above: **106/106 pass**, including built-CLI
  browser success/cancellation, exact grant schemas, bounded diagnostic capture
  and default payload omission.
- `pnpm --filter @omnius/execution exec vitest run tests/rapid-ui-core.test.ts
  tests/rapid-ui-visual.test.ts tests/rapid-ui-evidence.test.ts
  tests/rapid-ui-persistence.test.ts tests/rapid-ui-trace.test.ts
  tests/rapid-ui-native.test.ts tests/rapid-ui-native-uia.test.ts
  tests/rapid-ui-browser.test.ts tests/playwright-browser-contract.test.ts
  tests/playwright-browser-cancellation.test.ts tests/playwright-browser-observation.test.ts
  tests/playwright-browser-download.test.ts tests/playwright-browser-install.test.ts`:
  **311 pass**, **124 explicitly skipped** (one Windows case and 123 opt-in live
  browser cases). The skips are not platform passes.
- `pnpm --filter @omnius/orchestrator exec vitest run
  tests/ollama-backend-pool.test.ts -t 'passes the full JSON Schema'`: **2 pass**,
  35 unrelated cases filtered. The existing native transport forwards the full
  response schema; no backend transport patch was necessary.
- Updated smoke harness: **21 Node + 13 Python checks pass**. Synthetic planner
  request/reply receipts now retain bounded content, hashes, shape rejection and
  backend-request time. The first failed attempt remains immutable.
- Parser regressions reject empty target identities before they can become
  wildcard matches; legitimate unnamed controls still match exactly. Generated
  schemas preserve exact argument presence and use conservative Unicode bounds.
  Runtime authority remains independent of model shape validity. `git diff
  --check` passes. No inference occurs in these checks.
- `node scripts/rapid-computer-use/validate_plan_schema.mjs`: **18/18** independent
  Python JSON Schema checks against the built runtime/parser, including non-BMP
  strings, exact unnamed controls and full/partial host grants. The script is
  retained for reproduction; it requires an already installed `jsonschema`.

Typed host-verification report repair:

- `pnpm -r build`: all eleven workspace packages pass.
- `pnpm --filter @omnius/orchestrator exec vitest run
  tests/terminal-task-report.test.ts tests/terminal-task-report-runner.test.ts
  tests/host-tool-session.test.ts tests/completionFinalization.test.ts`:
  **99/99 pass**. Bound host receipts render as independent checks without
  invented command/exit details; false, malformed or unbound receipts remain
  failed/unverified. Later mutations mark the check stale, and UI checks cannot
  satisfy verification coverage for changed files.
- `pnpm --filter omnius exec vitest run tests/computer-use-host-session.test.ts
  tests/telegram-terminal-delivery.test.ts tests/telegram-admin-full-final-delivery.test.ts
  tests/telegram-conversation-evidence.test.ts tests/telegram-pdf-objective-isolation.test.ts
  tests/diagnostic-boundary.test.ts`: **129 pass, 10 fail**. The first three suites
  contribute 113 passing host/report delivery checks. All ten failures reproduce
  with the committed `d1343ba9` report implementation: both comparison runs have
  16 passes and identical ten failure names/messages, before report generation.
  The [baseline comparison](../../../computer-use/evidence/report-projection-baseline-comparison-20261012.json)
  records the source hashes, exact commands and retained private raw outputs.
  These pre-existing admission/fixture failures are not claimed as passes.
- `scripts/rapid-computer-use/replay_terminal_report.mjs` replays the retained
  successful run's ledger and committed finalization against the built renderer.
  The [separate replay result](../../../computer-use/evidence/host-terminal-report-replay-20261012.json)
  has one passed host check, no limits/next steps, and unchanged original artifact
  hashes. This is an offline display repair, not another live model run or a
  change to terminal authority. `git diff --check` passes.

## Visible three-model integration preparation

The user requested the GUI browser and small decision model driving the rapid
action space. This slice adds owned browser capture/input, a configured pinned
Photon client, full finite-policy task context and receipts, and bounded review
of an already committed completed task. The four-action fixture uses semantic
Title/Ready/Preview controls followed by an Apply canvas target and Cancel decoy.
The explicit headed harness admits one planner request, four actual policy
decisions and one point request, with no replanning or rules fallback.

Offline verification before any new CUDA launch:

- `pnpm -r build`: all eleven workspace packages pass.
- Execution browser visual helper, visual guard and geometry suites: **75/75**.
  Execution policy-context and core suites: **173/173**. Execution typecheck passes.
- CLI configured-point, model-preflight and command suites: **76/76**.
- Existing CLI host-session, command, compiled browser, index and delivery-receipt
  suites: **107/107**, including real CPU browser completion and SIGINT cleanup.
- CLI visual-browser suite: **11/11**, including eight review/Stop checks and
  three real owned-browser cases. Fake local model transports exercise the actual
  clients and one screenshot-derived point. Semantic steps make no vision call;
  missing authority makes zero point requests; the Cancel decoy cannot complete
  the task. These are transport/authority regressions, not model-quality results.
- `node --test scripts/rapid-computer-use/headed_host_session_smoke.test.mjs`:
  **26/26**, including exact call ordering, shared growth transitions, partial
  response cleanup and the real 40-character Kev/Photon revision format.
- `python3 -m unittest discover -s scripts/rapid-computer-use
  -p test_integrated_models.py -v`: **24/24**. The real broker validator remains
  strict for inference; initial readiness attests only its own pending token,
  owner and scope. Tests reject competing scopes and partial Qwen offload, retain
  exited runner identities until NVML allocation absence, escalate owned CPU
  children from TERM to KILL and preserve an unrelated process. Node syntax,
  Python compilation and whitespace checks pass.

The prepared live profile uses pinned Qwen, Kev and Moondream services on a single
supervised broker scope. Startup, actual placement and inference remain gated;
offline completion does not close the live GUI integration gate in the README.

Subsequent live verification closes that controlled-fixture gate: a fresh headed
run completed all four actions using one Qwen plan, four actual Kev selections
and one Moondream point. Independent read-only audits verified all **21 input
hashes**, **six captured receipt hashes**, the **four-event trace chain**, model
choice/action identity, point/click geometry, committed terminal authority and
process/allocation/lease cleanup. The final screenshot was inspected. The
69-artifact manifest is `computer-use/evidence/visible-host-attempt2-manifest-20261012.json`.
Original policy/point payload bytes and the pre-click PNG were deliberately not
retained; receipt hashes agree, but those raw payload hashes cannot be recomputed.
The screenshot is the visible completed state. WO-04 records measurements and
retains the original startup failure separately.

## Delivered commits

- `f7e432a4` — verified visible three-model task, screenshot, model/trace/terminal
  evidence and work-order update; pushed after independent evidence checks and
  owned browser/GPU cleanup. The two captured harness logs are explicitly tracked
  despite the repository's general log ignore rule, completing the 69-artifact
  manifest. Every listed artifact must be present in Git with its exact byte hash.

- `39564c55` — preserve the pinned Photon dependency overlay in supervised
  launches; pushed after 25 supervisor checks and a metadata-only 65/65 dependency
  check. Retains eighteen failed-startup artifacts and verified owned cleanup.

- `b7c7f88d` — owned headed browser, configured visual fallback, finite-policy
  context/receipts and shared model supervisor; pushed after the eleven-package
  build and visible integration preparation checks above. Its first live startup
  stopped before inference; failure and verified cleanup are recorded in WO-04.

- `3223c6a2` — render committed typed host checks accurately in shared terminal
  reports; pushed after the full build, 99 report checks, 113 host/delivery checks,
  baseline comparison and immutable-evidence replay described above.

- `1a490b44` — trusted negative-search receipts and bounded native recovery;
  fetched and pushed to `origin/main` after 127 targeted checks and dependency build.
- `291f8373` — exact native-save verification, geometry and desktop session lock;
  fetched and pushed to `origin/main` after 90 checks and execution typecheck.
- `f9df0362` — finite controller, policy/planner/workflow APIs and visual guard;
  fetched and pushed to `origin/main` after 63 checks and the workspace build.
- `406f4dcd` — managed browser and native semantic adapters, tools and registration;
  fetched and pushed to `origin/main` after the adapter, CLI and orchestrator
  checks above and recursive workspace build.
- `dc047ccf` — preserve stateful action admission, settle native effect receipts,
  bound stalled decision responses, and require native planner identity; pushed
  after 164 core/visual checks, the real runner regression and workspace build.
- `0e05a77a` — explicit verified Firefox BiDi transport and download limits;
  pushed after the follow-up checks above and workspace build.
- `7dc396da` — preserve unknown effects during concurrent duplicate dispatch;
  pushed after 175 core/visual/effect checks and execution typecheck.
- `f99d1326` — persist verified workflows and bounded redacted traces with no-input
  policy playback; pushed after 232 checks and the full workspace build.
- `255fcea3` — source/model pins, frozen finite and capture fixtures, guarded
  evaluation clients, model/runtime documentation and measured Qwen evidence;
  pushed after 69 offline Python checks and the earlier integrated verification.
- `2a32bd47` — staged Laya service and secret-free scoped GPU growth coordination;
  pushed after 35 service/lease contract checks and Python compilation.
- `2634ffa3` — measured Laya policy results and independently verified cleanup.
- `072c4f15` — pinned staged Kev service, verified offline before CUDA use.
- `f6c112c7` — bounded process-scoped GPU memory sampler and offline regressions.
- `0ec4a5b0` — pinned Photon service with separately gated raw residency and
  engine warmup, dependency manifest and 111 offline checks.
- `34d1e828` — measured Kev choices, sampled memory and verified cleanup.
- `676295e3` — retained Photon startup failure, raw placement and clean release.
- `cab9d2f0` — isolated compatible CUDA runtime, exact mapped-library checks and
  119 offline service/harness checks.
- `8452f548` — measured Moondream 20/20 synthetic grounding results, process-bound
  sampled memory, exact runtime records and independently verified clean release.
- `d6d639df` — bounded local computer-use command, runner ownership/terminal
  verification, configured planner and isolated browser example; pushed after
  477 checks and the full workspace build.
- `0f72246e` — one-request GPU growth bridge and owned-process smoke launcher;
  pushed after 30 offline checks and syntax/whitespace validation.
- `ea275cd3` — first live planner rejection, incomplete-ledger agreement and
  independently verified browser/model/lease cleanup, with 17 checked artifacts.
- `0a9ff3b3` — exact planner response schemas, bounded optional diagnostics,
  natural progress and empty-target safeguards; pushed after 419 affected checks,
  the independent schema checks and full workspace build.
- `d1343ba9` — verified live Qwen browser task, 29 hash-bound artifacts, independent
  request count, six progress messages, terminal authority and clean browser/GPU
  release; the first failed attempt remains unchanged.
