Skip to content

Verify browser interaction outcomes and reduce observation overhead - #48

Merged
revitteth merged 32 commits into
mainfrom
feat/context-usage-and-latency
Oct 1, 2026
Merged

revitteth merged 32 commits into
mainfrom
feat/context-usage-and-latency

Conversation

@revitteth

@revitteth revitteth commented Oct 1, 2026 •

Copy link
Copy Markdown
Contributor

Browser actions could acknowledge a click while leaving the application unchanged, and useful dialog controls could disappear behind background controls in the default snapshot. This change prioritises the active dialog, discovers styled native checkboxes, delivers trusted mouse or touch input, and refuses inactive profile-tab clicks with an explicit focus remedy. Input with an uncertain acknowledgement is not replayed automatically; callers must verify the application outcome.

The same release adds scoped usage and phase telemetry, bounded compact tab discovery, reader evidence improvements, and shorter observation paths. Passive batch/plan steps no longer consume human-action pacing, and waits enforce their host deadline.

Validation: local trusted-only dialog fixtures reproduce the old acknowledgement/effect mismatch and verify exactly one effect after the fix. The conformance matrix passes 24 seeded flows across desktop/touch, headed/headless and direct/extension control with identical semantic snapshots and expected final states. Controlled pacing and extraction measurements, their boundaries, and live workflow findings are in docs/next-release-research-2026-10.md and docs/latency-and-interaction-followup-2026-10.md. Transport latency remains different; these results do not establish human-speed navigation or complete host-model token accounting.

task check passed at 0b1997a: lint, extension guards, 61 Python tests, full Go suite, functional browser checks, secret/hygiene scans and race packages. Functional coverage: 20 passed, three explicit skips (two opt-in public-network flows and one extension-only tab-group flow in the direct lane); artifact and recipe smoke checks also passed.

Installed candidate 0.19.0-improve-0b1997a is verified on all three local daemons, both profile extensions report 0.7.11, and all three installed fixtures reach exactly one trusted effect with the dialog closed. A real Google Chat setup check exposed the dialog controls in the default frontier, toggled the styled external-members checkbox from false to true, and verified Cancel closed the dialog. No space or invitation was created. The owned tab and temporary route were removed.

The scoped Maix catalogue-administration companion is tracked separately in Don-Works/maix#197. The brw changelog prepares 0.20.0; no public release is claimed until its tagged artifacts finish.

Record exact HTTP body and MCP arguments/result boundary bytes, Unicode text counters, explicit character-based token estimates and separate binary/structured result counts. Forward metadata-only MCP and CLI projections into the canonical daemon ledger through a strict guarded endpoint without recursively logging ingestion. Preserve request correlation across proxy operations; reporting is bounded independently of operation cancellation.
Exercise numeric and JSON shape boundaries, bounded log coverage, concurrent rotation and recipe serialization with local fuzz properties.
Preserve legacy tab arrays by default; project full tab identities and existing lease states into text-only compact responses with explicit coverage. Filter before limiting, validate before enumeration, and document reuse of known tab IDs.
@revitteth revitteth changed the title Measure browser context costs and reduce observation latency Verify browser interaction outcomes and reduce observation overhead Oct 1, 2026
@revitteth
revitteth marked this pull request as ready for review October 1, 2026 17:57
@revitteth
revitteth merged commit 65b848b into main Oct 1, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant