Repository navigation
fix(cli): align MCP tool calls with compact browser tools - #1311
Closed
kingdomad (kingdomad) wants to merge 2 commits into
Closed
kingdomad (kingdomad) wants to merge 2 commits into
kingdomad (kingdomad) wants to merge 2 commits into
Conversation
The BrowserOS server no longer registers legacy browser tool names, so update the CLI to target the compact tool surface and argument shapes. Commands without a current MCP equivalent now fail explicitly instead of calling missing tools.
Contributor
|
PR author is not in the allowed authors list. |
Contributor
|
All contributors have signed the CLA. Thank you! |
Author
|
I have read the CLA Document and I hereby sign the CLA |
# Conflicts: # packages/browseros-agent/apps/cli/README.md # packages/browseros-agent/apps/cli/cmd/bookmark.go # packages/browseros-agent/apps/cli/cmd/click.go # packages/browseros-agent/apps/cli/cmd/dialog.go # packages/browseros-agent/apps/cli/cmd/dom.go # packages/browseros-agent/apps/cli/cmd/eval.go # packages/browseros-agent/apps/cli/cmd/file_actions.go # packages/browseros-agent/apps/cli/cmd/fill.go # packages/browseros-agent/apps/cli/cmd/history.go # packages/browseros-agent/apps/cli/cmd/info.go # packages/browseros-agent/apps/cli/cmd/interact.go # packages/browseros-agent/apps/cli/cmd/nav.go # packages/browseros-agent/apps/cli/cmd/open.go # packages/browseros-agent/apps/cli/cmd/pages.go # packages/browseros-agent/apps/cli/cmd/screenshot.go # packages/browseros-agent/apps/cli/cmd/scroll.go # packages/browseros-agent/apps/cli/cmd/snap.go # packages/browseros-agent/apps/cli/cmd/text.go # packages/browseros-agent/apps/cli/cmd/wait.go # packages/browseros-agent/apps/cli/cmd/window.go # packages/browseros-agent/apps/cli/mcp/client.go # packages/browseros-agent/apps/cli/output/printer.go
Vasilev Dmitrii (gHashTag)
pushed a commit
to gHashTag/BrowserOS
that referenced
this pull request
Sep 4, 2026
The swarm sat at zero bees of four while the tick refused with "nothing to choose" against 22 candidates and reported `claimed: 11`. stateOfDispatch had no case for 'failed'. It fell through to the bottom and returned 'awaitingReview', which QueenDelegationPolicy.claimOnIssue counts as a LIVE claim. So a dispatch set to 'failed' - the state the policy's own comment calls the one that most obviously means 'do this again' - held its issue instead of releasing it. The function could already PRODUCE that state two lines below, for a send-back that outlived its idle floor, and did not recognise what it produced. Measured 2026-09-04: five dispatches were set to 'failed' to return their issues to the pool (browseros-ai#1133, browseros-ai#1175, browseros-ai#1216, browseros-ai#1240, browseros-ai#1311). All five stayed in 'claimed' across four consecutive rounds. Every worker slot idle, fourteen real deficits the author could not file against because each already had an issue - one of the held ones. Work existed and was locked by the row recording its release. 'cancelled' is included for the same reason: it is a dead state and the registry treats it as free. Six new assertions in send-back-lease.test.ts, including the three that must NOT be swept up by it (accept, sendBack, and an unfinished dispatch), and one proving the case ignores the lease entirely - a failure is a failure at zero idle and above the ceiling, because nothing about it waits for a clock. 173 pass, 0 fail across the queen suite.
Vasilev Dmitrii (gHashTag)
added a commit
to gHashTag/BrowserOS
that referenced
this pull request
Sep 4, 2026
…e cause was fixed (#111) * feat(loop): publish the instruments the cron has been running all night These have existed only on one laptop. `git status` on the supervisor branch showed every one of them as untracked, which is the same defect this loop keeps finding in the swarm it supervises: work that exists and cannot be seen. 63 bee branches and 1321 files once sat in a container with 2 pushed. Twenty-one files, no dependencies, no network in the calibration run: heal.mjs eleven steps in a fixed order; four act, seven read loop.mjs state, the age-based lock, the append-only ledger, the dashboard reap lease free the volume, then the path fences push-work close-done make finished work visible, then clear it from the pool stale-escalations retire an escalation whose stated cause no longer reproduces clocks fields two defect classes that each cost an outage, checked every round fp-check coverage the checkers, run against material the WORLD calls good verdict-audit what the swarm claims, against what it actually pushed judge-packet assemble the unauditable for a judge; assembles, never judges brief-gate author refuse an unworkable brief; file from a measured deficit snapshot trend the readings, and whether a series has enough points to slope selftest 43 cases, every one proving the negative first Three rules are written into them and into the README: A checker never shown FAILING has not been tested. Six false accusations shipped in one night while synthetic fixtures agreed with the checkers that wrote them. Unreadable is not clean. A step that cannot reach its evidence reports ??, never ok - which is the only reason the first stale-escalation run did not retire six escalations on a probe that had failed to compile. A clock may only release what a clock can settle. Runtime is gitignored: state.json, ledger.jsonl, the lock, packets/, state/ and the rendered dashboards are produced by a run and mean nothing off the machine that produced them. No credential appears in any file; every database reference is process.env.DATABASE_URL, read inside the container. * fix(queen): a verdict that says failed was read as awaiting review The swarm sat at zero bees of four while the tick refused with "nothing to choose" against 22 candidates and reported `claimed: 11`. stateOfDispatch had no case for 'failed'. It fell through to the bottom and returned 'awaitingReview', which QueenDelegationPolicy.claimOnIssue counts as a LIVE claim. So a dispatch set to 'failed' - the state the policy's own comment calls the one that most obviously means 'do this again' - held its issue instead of releasing it. The function could already PRODUCE that state two lines below, for a send-back that outlived its idle floor, and did not recognise what it produced. Measured 2026-09-04: five dispatches were set to 'failed' to return their issues to the pool (browseros-ai#1133, browseros-ai#1175, browseros-ai#1216, browseros-ai#1240, browseros-ai#1311). All five stayed in 'claimed' across four consecutive rounds. Every worker slot idle, fourteen real deficits the author could not file against because each already had an issue - one of the held ones. Work existed and was locked by the row recording its release. 'cancelled' is included for the same reason: it is a dead state and the registry treats it as free. Six new assertions in send-back-lease.test.ts, including the three that must NOT be swept up by it (accept, sendBack, and an unfinished dispatch), and one proving the case ignores the lease entirely - a failure is a failure at zero idle and above the ceiling, because nothing about it waits for a clock. 173 pass, 0 fail across the queen suite. --------- Co-authored-by: Dmitrii Vasilev <trackgmedernj@hotmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR Description
Summary
Updates the BrowserOS CLI to call the MCP server's current compact browser tool surface instead of legacy tool names that are no longer registered. Commands that no longer have a server-side equivalent now fail explicitly with a clear unsupported message instead of attempting a missing MCP call.
Main changes
tabs,navigate,act,snapshot,read,grep,screenshot,pdf,download,tab_groups,windows,wait,evaluate, andupload.tabstool'spages[]structured output.CallToolcalls.tabsoutput.Risk / Review focus
e12; reviewers should check compatibility with existing CLI usage expectations.dragandfocusremain unsupported because the compact MCP surface does not expose equivalent element-based behavior.Validation