Browser troubleshooting

The Synara browser is a shared Chromium surface scoped to the owning task. Browser failures are safest to handle with an observe–act–observe loop: inspect the current page, perform one bounded action, then inspect the resulting state.

Browser is unavailable

Start with browser_status when the provider has browser tools.

Check:

• The task owns or can acquire a browser tab
• The desktop browser host is running
• The task is still authorized to use the browser
• The tab was not closed or crashed
• The operation targets the current task’s browser scope

Opening another chat does not necessarily destroy the task’s browser runtime. Do not create repeated tabs simply because the browser surface is hidden.

No tab is assigned

Use browser_tabs to inspect tabs in the task scope, then browser_open to create or reuse one.

• Omit tabId to use the provider session’s assigned tab
• Use only tab IDs returned in the same thread scope
• Use reuse:false only when a genuinely new tab is required
• Use show:false when background work is appropriate and the page does not require human interaction

Do not guess a tab ID from another task.

Snapshot reference is stale

Element references are bound to the snapshot that produced them.

Correct target shape:

{
  "target": {
    "ref": "e3",
    "snapshotId": "current-snapshot-id"
  }
}

A bare e3 can never safely identify an element after navigation, resize, scrolling, or page mutation.

When a stale-reference error occurs:

1. Take a new browser_snapshot.
2. Find the element again.
3. Use the new ref and snapshotId together.
4. Retry only the intended action.

Action timed out

A timeout does not prove that the action failed.

For clicks, typing, navigation, form submission, uploads, or evaluation:

1. Do not immediately repeat the mutation.
2. Take a fresh snapshot.
3. Check the URL, visible text, form value, and page state.
4. Use browser_logs when console or network evidence matters.
5. Retry only when the observed state proves the effect did not commit.

Repeated blind retries can submit forms twice, create duplicate records, or overwrite text.

Typing produced no visible text

Check:

• The target is editable and enabled
• The snapshot is current
• The field was replaced when you expected append behavior, or appended when you expected replacement
• A re-render did not replace the input during the action
• The page did not reject the value after an input or change event

Re-snapshot and inspect the field value. For long text, use bounded chunks and verify after each committed step rather than assuming one large action succeeded.

Click opened an OAuth or permission popup

Stop browser automation when the result reports that human action is required.

1. Reveal the owning task’s browser surface.
2. Let the user complete sign-in or consent in the visible popup.
3. Do not inspect, type, or copy credentials through browser evaluation.
4. Resume only after the user confirms completion.
5. Take a new snapshot because the original page state may have changed.

Navigation succeeded but the page is not ready

Wait for a bounded condition rather than a fixed long sleep when possible.

Example:

{
  "conditions": [
    {
      "kind": "text",
      "text": "Dashboard",
      "state": "present"
    }
  ],
  "timeoutMs": 15000
}

After the wait completes, take another snapshot. A load event can finish before client-side content is usable.

Browser logs are empty

browser_logs returns bounded console, exception, and network metadata for the exact tab. It does not return every header, request body, or response body.

Empty results can mean:

• No matching event was captured
• The relevant event occurred before capture
• The failure is outside the page
• The evidence was truncated or not retained
• The wrong tab was inspected

Do not treat an empty log result as proof that no browser or network failure occurred.

Upload was rejected

Browser uploads accept regular files inside the task workspace.

Check:

• The path is workspace-relative
• The target is an enabled input[type=file]
• The path does not traverse outside the canonical workspace
• The file is not a directory
• A symlink does not escape the workspace
• The upload matches explicit user intent

Never upload credentials, private keys, environment files, or proprietary artifacts without explicit authorization.

Keyboard command was rejected

The browser accepts page-scoped key chords and blocks privileged OS, application, browser, and clipboard shortcuts.

Use normalized keys such as:

{ "keys": ["Control+A", "Backspace"] }

Use visible browser controls instead of attempting to bypass policy with privileged shortcuts.

Screenshot differs from semantic snapshot

A screenshot answers pixel-level questions. A semantic snapshot answers structure, visible text, roles, state, and actionable references.

Use:

• browser_snapshot for interaction and accessibility structure
• browser_screenshot for layout, clipping, colors, canvas, or image evidence
• Both when visual appearance and interactive semantics disagree

Do not interact from screenshot coordinates when a current semantic target is available.

Evaluation failed or changed the page

browser_evaluate runs a bounded main-world expression and can mutate the page.

Before retrying:

1. Inspect the current page.
2. Determine whether the expression partially committed.
3. Prefer normal browser actions for the recovery.
4. Avoid using evaluation to bypass navigation, network, upload, or native-surface policy.

Browser crashed or disconnected

Record:

• Task and tab ID
• URL or annotation ID
• Last successful tool call
• Exact error code
• Whether the action may have committed
• Whether the tab appears in browser_tabs
• Whether a new tab reproduces the failure

Close a crashed tab only after preserving evidence. browser_close is permanent for that tab and invalidates every reference.

Browser recovery checklist

• The correct task and tab scope were confirmed
• A current snapshot was taken
• Every element action used its matching snapshotId
• Ambiguous mutations were observed before retrying
• OAuth and credential steps were left to the user
• Logs and screenshots were used only when relevant
• Upload paths stayed inside the task workspace
• The final user-visible result was verified

Continue with Browser verification for the normal workflow or Diagnostics when the browser host and task state disagree.