Repository navigation
Conversation
fix: decode and stop silently truncating inbound integration messages
Provide LLM reasoning effort param
supabase integration
Signed-off-by: ahmad-ajmal <ahmadajmal1514@gmail.com>
fix tunnel failed silently and get error from API rather than log file
Signed-off-by: ahmad-ajmal <ahmadajmal1514@gmail.com>
Signed-off-by: ahmad-ajmal <ahmadajmal1514@gmail.com>
Squash of the original korivi-Exp1 prototype (web agent V1/V2, password vault, onboarding error wrap) applied on top of current main. Conflicts resolved against main's newer browser adapter and sidebar; the local-only slow_mode and profile-interview workarounds were dropped. The page still targets the old task API; it is adapted in the next commits.
A real Chromium that agents drive and the user watches live, usable by every kind of agent (main and chat sessions, a dedicated Mini Browser chat, scheduled runs, Agent App trigger runs and the new browser_agent sub-agent). Engine (app/mini_browser): - dedicated "mini-browser" thread + event loop owns Playwright; callers on any loop marshal in (fixes sub-agents, which run on their own loops) - one tab per agent session with per-tab locks, owner labels, take-control, follow-view, crash and driver-death recovery, idle shutdown, profile lock - live view via CDP screencast of the viewed tab, only while a UI tab is subscribed, latest-frame-wins delivery - ad blocking via CDP Network.setBlockedURLs (context.route disabled the HTTP cache); CraftBot's own UI origins blocked browser-wide - human-like mouse paths, typing rhythm and scrolling - 13 mini_browser_* actions; every result carries a fresh page observation Password vault: Fernet + Windows DPAPI in .credentials/, atomic writes, no silent data loss, strict origin matching; autofill only for the current page's origin and never exposed to the model, logs or UI. Platform: mini_browser_* WebSocket protocol and lanes, new session type mini_browser, settings section, error codes, AgentBase run-state / stop / delete / sub-agent / shutdown hooks, skill, prompt rule and AGENT.md. Frontend: Mini Browser page (sidebar entry, live view with IME/paste/ shortcuts, tabs, take control, passwords dialog), 7 languages.
- Agent input never lands after the user takes control. Waits for an element run in 250 ms slices that check control; clicks press through page.mouse at a point just checked to hit the element instead of Playwright's click, which kept waiting inside the driver. Focus, clear, caret, select, upload and whole-value inputs, and the vault login's fill and Enter, act on element handles right after the control check. - Agents cannot close a tab the user controls. - Opening a tab reserves its slot before the first await, so concurrent opens never exceed the browser-wide or per-owner tab limits. - Agent screenshots cover password and secret fields, including fields that were password fields before a "show password" toggle. - A saved password in a download's suggested name is masked before it becomes the file name, a log line or a notice. - Cut deletes the selection only after its own copy reached the clipboard, in the tab it was copied from (copy replies echo requestId). - One settings normalizer (mini_browser.config.clean_setting) for app.config and MiniBrowserSettings. - Browser tests skip only when Playwright or Chromium is not installed; CRAFTBOT_REQUIRE_BROWSER_TESTS=1 makes that a failure too.
Signed-off-by: ahmad-ajmal <ahmadajmal1514@gmail.com>
…ance fix #475: separate Windows cmd and PowerShell actions
fix: prevent integration settings loader from hanging
…ns-cleanup Make agent aware of NOTE and make missions obsolete
…replies fix: enable Telegram replies, owner control, reply context and typing feedback
grep improvement update using ripgrep
Threads (and other sites that link addresses as you type) put agents into
a delete-and-retype loop, with link previews flipping between sites:
- Web and e-mail addresses are inserted whole, like a paste. Typed key by
key, the site saw every partial address ("https://e", ...,
"example.co") and previewed the wrong ones.
- After typing into a rich text box (contenteditable), the box is read
back and the result says whether it holds the whole text; links, link
previews and styled mentions are named as the site's formatting.
- The same text is typed into the same page at most 3 times in 10
minutes; the 4th attempt is refused with a "stop and check or ask the
user" message. Only a hash of the text is kept.
- The mini-browser skill and the type action describe link previews and
say never to delete and retype a post because it looks different.
Sound: Playwright starts Chromium with --mute-audio, so no page could play
sound. New setting mini_browser.sound (default true) drops that switch; it
applies at the browser's next start.
The compact page observation showed only the first 40 characters of a post or message box, so a post with a link often appeared as "https://example.c…": a broken-looking link the agent took for its own typo, one more trigger of the Threads retype loop. - Textareas and rich text boxes show up to 300 characters, cut between words (also in the tighter fallback format), and say how long the full text is when it is cut. - The read-back after typing covers the whole editor, also when the element is one of its paragraphs. - The retype guard counts a retry that only writes an address differently ("https://" added or dropped) as the same text. Checked against the real Lexical editor with its AutoLinkPlugin: typing an address key by key changed the link five times (example.c -> example.co -> example.com -> ...), now it is created once. Sound was measured in the Windows mixer: Chromium plays (live-view and agent clicks) with sound on, and is silent with it off.
- Dropdowns: Playwright's select_option keeps waiting in the driver while the option is missing (force does not stop that retry), so a choice could land after the user took control. Each attempt is now one in-page step that finds the option again and selects it only if it is there and enabled (input + change, like Playwright); between attempts the wait only reads, and every attempt checks control first. - Closing the last tab: the owner and take-control checks run again after the replacement tab has opened, right before the close. - Cut: the copy is marked as a cut and the backend keeps the text it copied (memory only, 30 s). The delete is a "cut" input event in the ordered input queue, after whatever the user did meanwhile, and runs only while the selection is still exactly that text; otherwise it is skipped, so nothing that is not on the clipboard is deleted.
Signed-off-by: ahmad-ajmal <ahmadajmal1514@gmail.com>
- Cut: only a complete copy can be cut. A selection longer than the copy limit, or one with a saved password masked out, reaches the clipboard only in part, so it is copied but never deleted; the delete compares the whole selection. Pending cuts are kept per request, with ids scoped to the viewer connection (and unique per view mount), so two viewers' cuts never share a record and a mismatched delete leaves the others intact. - Retype guard: scoped to the same field (data-mb-field) and to text that was not sent; Enter or the page moving on resets it. Addresses only ignore "https://" and "www.", so different links are different texts. Filling several fields with the same value is never refused. - Navigation: control is checked again right before an operation runs (after crash recovery and bringing its window forward), right before navigate's goto / back / forward / reload, and before a new tab opened with a URL loads it.
Fixed action set accumulation issue causing system prompt to bloat
…gnment-v1.4.4 fix: change google scopes
feat #479: add interactive Generative UI outputs
PR #483 was merged into main by mistake; this merges it into V1.4.4, its intended base. V1.4.4 made a session's <actions> a function of its type and turned add_action_sets into a read that delivers schemas into the event stream (SessionManager.add/remove/get_action_sets, Session.action_sets and the create_session action_sets argument are gone). The PR was written against the old model, so it is ported here: - SessionType.BASE_ACTION_SETS gives mini_browser sessions the mini_browser set; ws.ensure_session no longer passes action_sets. - load_action_sets keeps the PR's lenient name matching, web_agent alias and unknown_sets hint; the PR's remove_action_sets changes are dropped with the API. - _apply_workflow_capabilities keeps the PR's user-skill action sets and delivers them through load_action_sets. - The router keeps the PR's prompt filter, without integration_essentials (removed on V1.4.4). - AGENT.md (both copies) and the heartbeat skill describe the V1.4.4 model; remove_action_sets references dropped. - Mini Browser tests updated to the V1.4.4 API. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.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
4 participants
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.
What and why?
Combining the generative UI and mini browser into a left panel for each session. Making our built-in browser session-based. Major UI update to make them display in the left panel without being overwhelming.