Frontier browser runtime: six user outcomes and installed-release acceptance

[shaun0927]

Objective

Deliver the six requested user outcomes in OpenChrome: faster verified tasks, reliable account/authentication reuse, efficient parallel work and memory, no pre-opened empty Chrome requirement, no unsolicited headed interference, and accurate headless observation with controlled human takeover.

Scope decision

Keep the host responsible for goals and LLM strategy. Reuse the existing CLI/MCP runtime and contracts. An optional browser observation client may consume the same authorized API; mandatory desktop apps/extensions and unrelated integrations are outside this implementation. Existing product direction #1359 remains relevant. Coordinate session manifest work with #1578.

Work breakdown

• [F01] test: establish reproducible browser task and resource baselines #1650 — F01 test: establish reproducible browser task and resource baselines
• [F02] refactor: define canonical task identity and lifecycle across runtime surfaces #1651 — F02 refactor: define canonical task identity and lifecycle across runtime surfaces
• [F03] fix: enforce lazy startup and explicit browser window visibility policy #1652 — F03 fix: enforce lazy startup and explicit browser window visibility policy
• [F04] feat: expose task-scoped browser observation with bounded capture overhead #1653 — F04 feat: expose task-scoped browser observation with bounded capture overhead
• [F05] feat: unify account selection and authentication state management #1654 — F05 feat: unify account selection and authentication state management
• [F06] feat: support exclusive human takeover and verified task resume #1655 — F06 feat: support exclusive human takeover and verified task resume
• [F07] fix: expose cancellation deadlines and uncertain action outcomes #1656 — F07 fix: expose cancellation deadlines and uncertain action outcomes
• [F08] perf: execute bounded action sequences with task-specific waits #1657 — F08 perf: execute bounded action sequences with task-specific waits
• [F09] fix: validate observation freshness and bound cache reuse #1658 — F09 fix: validate observation freshness and bound cache reuse
• [F10] feat: schedule browser work across tabs accounts and shared resources #1659 — F10 feat: schedule browser work across tabs accounts and shared resources
• [F11] feat: add Windows browser process accounting and resource admission control #1660 — F11 feat: add Windows browser process accounting and resource admission control
• [F12] fix: isolate client tool disclosure and lifecycle semantics across transports #1661 — F12 fix: isolate client tool disclosure and lifecycle semantics across transports
• [F13] release: validate installed candidate and upgrade path against browser acceptance scenarios #1662 — F13 release: validate installed candidate and upgrade path against browser acceptance scenarios

Outcome gates

• Speed: F01/F08/F09 demonstrate reduced time/cost with matched task correctness; failed attempts included.
• Authentication: F05/F06/F07 verify intended account, coalesced refresh and resume without duplicate actions.
• Parallel/memory: F10/F11/F12 demonstrate bounded queues, fair work, Windows measurement validity and owned-resource reclamation.
• Empty Chrome: F03 keeps initialization browser-free; F11 reclaims eligible idle resources.
• No interruption: F03/F06 preserve the selected visibility policy through recovery/takeover.
• Observation: F02/F04/F06 show the correct task/target, freshness, wait reason and exclusive user control.
• Installed release: F13 verifies the actual tarball, upgrade and newly published exact package.

Integrated acceptance

On Windows, keep another application foreground while four independent authenticated tasks execute. No managed Chrome exists before browser demand, and background policy causes no unsolicited visible window/focus changes. Expire one task's session: only that task/account pauses, the observer identifies its exact target/reason, the user gains exclusive control, and validated account/access permit resumption. Independent work continues. Reconnect one client and retrieve the same status/results. Verify results from fixture server state/downloaded files, then confirm owned-resource reclamation under the declared idle policy.

Evidence and closure

Baseline source: afaf391; published comparison: openchrome-mcp 1.12.9 (npm gitHead 25dd4f8). Pin every candidate by SHA/package integrity. Unit, real-Chrome fixture, external-account and installed-release results are separate. Unknown telemetry or absent test credentials is pending/inconclusive, never success. Record commands, environment, repetitions, sample-validity policy and independent outcome oracles in durable CI/issue-accessible evidence. Do not close this parent on source-only checks or self-reported agent success.

Execution order

F01 first; F02/F03 establish contracts; F04/F05/F07 enable F06; F08/F09 improve bounded execution; F10/F11/F12 validate scaling and multi-client behavior. Start F13 installed-package smoke early and complete it after all dependent gates pass.

[shaun0927]

All 13 detailed implementation issues have been opened (#1650–#1662). Foundation implementation is in draft PR #1664, head 7d6bc656ead81cd8449f6f2070d287af16684346.

Implemented and tested in that PR: protection for parallel in-flight tabs, bounded target/capture queues, deadline/cancellation admission for interact, session/lane ownership for observation, per-client disclosure, suppression of automatic headed fallback under headless configuration, and Windows managed-Chrome process-tree monitoring through lazy connection/shutdown.

Verification: 266 distinct local focused regressions passed. Build/lint/structure/tier/schema/capability checks passed. GitHub independently packed and installed the candidate on Windows and Linux; both passed 25 fixture receipts, seven valid memory samples, lazy startup and owned Chrome exit. Exact package hashes and tested revision are recorded at #1664 (comment) . Published 1.12.9 also passed the final fixture rerun after earlier intermittent failures; no comparative speed or failure-rate claim is made.

This is not completion of the full roadmap. Canonical task lifecycle, account-aware authentication refresh, exclusive human takeover, guarded sequences, robust observation freshness, account/resource scheduling, full transport/client matrix, long-duration/agent benchmarks, real-account acceptance and new published-release verification remain open. Each child issue now distinguishes the implemented foundation from its outstanding acceptance. The PR has not been merged and no npm release has been published.

[shaun0927]

Additional implementation and validation landed on draft PR #1664 at 47da230802f68625a55322689535db3cb4d3ddc5: explicit/uncertain storage restoration, preservation of failed snapshots, non-destructive tab-capacity admission, headed-registration failure cleanup, and outside-checkout installed-package tests.

Results, exact revision/artifact information, limitations and remaining work. Ordinary CI and installed Windows/Linux acceptance passed. Locally, 89 distinct focused assertions and both installed fixture runs passed; optional full suite was not run. These results do not complete the roadmap or demonstrate external-site authentication or comparative speed.

All 13 child issues remain open. Required implementation still includes canonical task state, the observer UI/timeline, account validation/shared refresh, exclusive human takeover/resume, complete cancellation propagation, bounded sequences, freshness-safe reuse, account/resource scheduling, pressure-based reclamation, and the full transport/client matrix. Comparative workloads, upgrade validation and a newly published exact-version smoke are also outstanding. This work is unmerged and unpublished.

[shaun0927]

The confirmed implementation scope supersedes the original thirteen-feature roadmap. The user explicitly asked to keep only capabilities that OpenChrome needs as a host-neutral MCP harness.

Confirmed work	Tracking	Implementation
Predictable lazy browser lifecycle and explicit visibility	#1652	Owned-target cleanup, headless policy, real foreground-event check
Reliable login/profile continuity and page/account validation	#1654	Per-context restore/watchdogs, preserve fresh state, explicit verify
Safe bounded parallel work and resource cleanup	#1660	Target reservations/queues, screenshot/operation bounds, no eviction of active work
Measured latency improvements	#1650	Bounded parallel tab metadata reads; repeatable per-process measurement
Honest cancellation/deadline/disconnect outcomes	#1656	Scoped cancellation, recovery deadline, raw CDP draining, no uncertain-write replay
Lightweight headless observation	#1653	Current/recent operation metadata plus existing page evidence
Minimal human pause/drain/resume	#1655	Input lease, account/page checks, reference invalidation, held-session protection
Common protocol/install/upgrade verification	#1661, #1662	Focused tests, installed package, Windows/Linux Chrome CI and controlled upgrade fixture

Implementation is in PR #1664, source commit 0ce93cb33e560c7d129c3c32ea5a4ef27cdbc3a8. It has 189 passing focused tests and green general plus Windows/Linux installed-browser CI. Detailed final measurements and upgrade evidence will be recorded with the PR results.

The original task-planner expansion (#1651), new sequence engine (#1657), speculative snapshot reuse (#1658) and general scheduler platform (#1659) are removed from the confirmed scope. Existing host planning, execution tools and per-target queues are reused. Closing those items as not planned is scope reduction, not a claim that their original broad designs were implemented.

No automatic CAPTCHA/MFA bypass, universal login guarantee, large desktop dashboard, new planner or blanket performance-superiority claim is part of this work. Remaining release, external-site and host-specific validation must be distinguished from the completed implementation and fixture checks.

[shaun0927]

Executed the confirmed implementation scope in source commit 0ce93cb33e560c7d129c3c32ea5a4ef27cdbc3a8.

• 190 focused assertions passed (189 across 17 suites, plus one concurrency/order regression).
• General CI passed: https://github.com/shaun0927/openchrome/actions/runs/34128708454 . Optional full-test skipped.
• Windows/Linux installed real-Chrome CI passed: https://github.com/shaun0927/openchrome/actions/runs/34128708444 . Tested merge 27dbf6fadf51bd436458b7c92b26077992cf82f6.
• Outside-checkout candidate: 2,383 dist files match the build; local Chrome 152.0.7977.82 passed lazy launch, capacity refusal, four isolated accounts, 25 server receipts with no duplicate writes, cancellation draining and verified human resume.
• Upgrade from candidate 47da230: all four saved accounts authenticated after reinstall/restart without revisiting login. Both old and final Chrome exited.
• Local installed/upgrade foreground observation was valid, with zero owned-Chrome foreground events.
• Performance results are mixed: baseline trial p50 11.95/11.95/10.02 ms; candidate 6.23/16.38/5.83 ms. Worst trial p95 increased 17.74→43.51 ms. No consistent/global speed improvement or competitor superiority is claimed.

Candidate remains unpublished version 1.12.9, SHA-512 CI/tJGtltdGA1VIvTUabgaop6owq0pgx9l/8VBU46t7dk3QCWkGZJWtx+lQvMQuwNEGX4w8XUGCljJ5qY3o0yg==. Full report: docs/2026-09-07_openchrome-runtime-report.md; usage: docs/agent/browser-control.md.

External-site SSO/MFA, real-person and host-specific usability, long-running memory stability, merge and public release are not proven or completed by these fixtures. Performance acceptance remains evidence-limited; the other required functional fixture checks passed.

[shaun0927]

Final-head verification is complete for 323148f3fcc2c546d8e3b12ae4de9d25d508e368.

• General CI passed: https://github.com/shaun0927/openchrome/actions/runs/34129728590
• Windows and Linux installed real-Chrome acceptance passed: https://github.com/shaun0927/openchrome/actions/runs/34129728709
• Optional full-test was skipped. Execution source is unchanged from the artifact-verified 0ce93cb; the final commit adds the result report, usage instructions and bounded-concurrency regression.
• Worktree is clean; PR is ready for review. Mixed performance results remain explicit, and no merge or npm publication has been performed.

[shaun0927]

PR #1664 is merged into main as 8a2434542e43eb34e3e887dccd0f5a2922c891a0. Release version 1.13.0 is committed on main at 3f6fd67f3e2ed245ba30374ec1e55b689391f79c.

GitHub release published with the tested tarball, checksum and local acceptance report: https://github.com/shaun0927/openchrome/releases/tag/v1.13.0 . General main CI and Windows/Linux installed Chrome CI passed: https://github.com/shaun0927/openchrome/actions/runs/34144951543 and https://github.com/shaun0927/openchrome/actions/runs/34144953859 .

npm publication remains blocked by absent authentication (npm whoami returns ENEEDAUTH). The user has been asked to log in locally; no credential values were requested in chat. npm 1.13.0 and the latest tag are not claimed published. Performance and external-site/endurance limitations remain unchanged.