Repository navigation
Release: staging → main (126 commits) - #807
Draft
mindsdb-release-train[bot] wants to merge 126 commits into
Draft
mindsdb-release-train[bot] wants to merge 126 commits into
mindsdb-release-train[bot] wants to merge 126 commits into
Conversation
One-off complete() calls (structured output, compaction, rule filter) never resend the same history, so a message mark there paid a cache write nobody read. The planning loop streams, so it keeps the mark. Also import SESSION_CONTEXT_MARKER at module level and update its comment now that the Anthropic transport caches the shared part too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n limit Turns that reach the per-request token limit had often had an answer for most of the turn, then spent the rest investigating a property of the data they never tested, and withheld the answer at the limit. - Once a turn has spent half of max_turn_tokens, a one-time note rides with that round's tool results: give the answer now if there is one, state what it assumes, test a surprising data property once, keep working only on a named error or on work still needed. It does not stop the turn. - The limit hand-back no longer reads as "withhold anything unexplained": a result the tools computed is given with its assumptions. Figures the tools never produced are still refused. - The verifier transcript drops anton's notes when they arrive as a text block beside tool results, as it already did for plain-string notes, so the verifier never reads the note as the user asking to wrap up. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The act-first discipline said to act on a reasonable reading but not how to choose between readings. It now asks for the most literal one. A general rule keeps a summary line or final answer from dropping values the question asked for. The default prompt snapshot digests are updated for the two new lines. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(session): internet_research_enabled flag for data-only sessions A session that must answer from the connected data only (evals such as DataAgentBench) could turn off the web tools, but the system prompt still told the agent it has full internet access through the scratchpad and the scratchpad's web_search() helper kept working. internet_research_enabled=False (ANTON_INTERNET_RESEARCH_ENABLED) now: - turns off both web tools, whatever their own flags say; - makes the scratchpad's web_search() raise before any model call; - swaps the prompt's internet guidance, the helper's mentions in the prompt, the scratchpad tool and two nudges, and the public-data-sources skill line for "work with the connected data only". It is a rule for the agent, not a network block. It is separate from the web tool flags because cloud_turn turns those off while its scratchpads keep the internet; cloud_turn doesn't set it, so cloud sessions are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(prompts): count the two no-internet nudge variants Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(generate_artifact): no web tools in the gathering loop without internet research With internet_research_enabled off, generate_artifact's own gathering loop still offered web_search and web_fetch, ran them through the session's fallback handlers, and fed the public-data-sources catalog to its fetch step. It now leaves both tools out of its tool array, shows the scratchpad schema without web_search(), refuses a web call the model names anyway, and reads no catalog. Also from review: the no-internet prompt drops the follow-up suggestions about public datasets, the manager warns once per process about a factory without the flag, and _build_core_tools reads the flag like the other call sites. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…hing Bugfix/anthropic byok prompt caching
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t without tools Measured on long analysis turns: the budget checkpoint got the answer given with its assumptions, the completion verifier read those assumptions as unfinished work, and the continuation went back to investigating until the token ceiling, ending with no answer. Past half of max_turn_tokens, an INCOMPLETE verdict now gets one reply without tools instead of a working continuation. The note carries the verifier's reason and asks the reply to add anything the results in hand contain, to say plainly what they do not show rather than give a value for it, and otherwise to restate the answer with its assumptions. The turn ends on that reply. Below the checkpoint nothing changes. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…versation Bugfix/pin memory prompt per conversation
…rocesses The workspace listing sits in the system prompt ahead of the whole conversation. A new root file or scratchpad changed it, so the next turn re-wrote the conversation to the prompt cache. It is now pinned with the memory text and labelled as the state when the conversation started. The pin lived only in the process, but every cloud turn runs in a new process, so cloud turns never reused it. It is now also saved beside the conversation and loaded when the process has no copy. The process copy still wins, and a saved part of the wrong type is rebuilt. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent code can write the saved pin, and a cloud turn always starts from it, so a forged memory or scratchpad text reached every later prompt. Only the workspace listing is saved now, as names that must be plain on load, and the host renders the text. Memory, scratchpad and read parts stay in the process as before. The pin is saved once per prompt build and cleared with an empty record. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-conversation Pin the workspace listing per conversation and save the pin for new p…
…ll fails (#724) * fix(minds_client): name the request, status and server message on a failed Minds call A failed publish reached the CLI and the agent as a bare "HTTP Error 401: Unauthorized": no endpoint, no server reason. minds_request now raises MindsHTTPError, an HTTPError that keeps the response body and the request line, so every reader of the body (the publisher's 409 check, test_llm, cowork-server's upload classification) still gets it. str(), reason and code are unchanged. Non-HTTP failures keep their type and gain the request line. describe_minds_request_failure turns that into one line for the publish tool result and the /publish and /unpublish messages; server text is markup-escaped before it reaches the console. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(minds_client): hint what to check, never a cause, on a failed request The failure line reused the desktop "common reasons" advice, which for any 5xx says "a temporary outage" — the story the agent invented in the first place, now handed to it, even when the server's own message says otherwise (a pip timeout, a KeyError). The line now ends with what to check, and a 5xx says only that nothing here shows the cause. Also covers /publish and /unpublish (list, remove) with tests, and stops each local test server waiting 0.5s on shutdown. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(publisher): a share-org refusal keeps its own instruction over the real transport Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(chat): show the failed request on a /publish 401 too The 401 branch cleared the key and gave credential advice but never said which request failed or what the server answered. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… an entry (#723) * fix(memory): keep hand-written lines when saving, editing or deleting an entry rules.md and lessons.md are files people edit by hand, and Cowork exposes them as project files. Saving a rule rebuilt rules.md from the parsed entries, which dropped typed notes, rules above the first heading and rules under any heading but Always/Never/When. Editing or deleting a lesson did the same to lessons.md. Saving a rule now inserts one line at the end of its section, and editing or deleting an entry changes only that entry's line. Entries with no metadata no longer gain "confidence:None source:None". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(memory): read and edit lessons from the same parse get_lessons stripped the whole file before parsing, while edits parsed it as written. A file whose first bullet was indented returned that lesson from get_lessons, but update_lesson and del_lesson could not find it. Both now parse the file as read, ignoring leading whitespace on the first non-blank line, so every lesson returned can be edited or deleted. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…746) A failed Minds request read "returned HTTP 401: <server text>. The API key was rejected: check ANTON_MINDS_API_KEY." with nothing marking where the server's words ended, so the agent told users the server had pointed them at ANTON_MINDS_API_KEY. The server's message is now quoted and the advice that follows is labelled as anton's. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…heckpoint Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The halfway note and the tool-less reply after an unfinished verdict were tuned on one benchmark of single-value questions. Past half the token limit they touch many requests doing legitimate long work, and the tool-less reply can cut off a part of the request that was never done. Keep the one change that holds for any request: at the limit, a result the tools computed is given with its assumptions instead of being withdrawn as a guess. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…them (#741) * feat(analytics): label PostHog events with the deployment that sent them The direct PostHog events now carry `environment`, read from ANTON_DEPLOYMENT_ENV and bounded to a short slug. It is "" where no host set it, so an unlabelled install reads as unknown rather than production, and a query can tell a build that carries the field from one that predates it. Refs ENG-3454 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(analytics): say every direct-to-PostHog event carries environment The label is stamped on all events in _POSTHOG_EVENTS, not only the two measurement events the page names. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Snyk has created this PR to upgrade react-dom from 19.3.0 to 19.3.1. See this package in npm: react-dom See this project in Snyk: https://app.snyk.io/org/mindsdb-default/project/8cb1873f-0d1e-4d36-bd82-9a6a798bcd00?utm_source=github&utm_medium=referral&page=upgrade-pr
Fix/token budget checkpoint
GitHub's runners take close to a second on this check, so the 1s bound failed on slower machines: three times in a row on anton#549, which doesn't touch the write check. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…urns (#721) * Stop asking Jev for the session after MindsHub refuses it MindsHub refuses Jev with a 403 to an organization with zero data retention on (or an admin's model rule), and would refuse every later check the same way. After the first refusal the completion check uses the LLM verifier alone for the rest of the session. Other failures (timeouts, 402, 5xx) keep asking, as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Honor the organization's Jev opt-out on cloud turns TurnRequestV1 gains jev_decisions. Only a JSON false turns the completion check's Jev call off; a missing field or any other value keeps it on, so a controller that does not forward it yet changes nothing. cowork-server sends it (mindsdb/cowork-server#761) and scratchpad-controller forwards it (mindsdb/scratchpad-controller#106). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Name the completion check, not the retired shadow probe, in Jev's comment Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Fail the Jev opt-out closed and close the client after a refusal A turn request's jev_decisions is on only when missing or true; any other value turns Jev off, since a malformed opt-out must not read as consent. After MindsHub's first 403 the session closes its Jev client, and the comment now says the latch lasts the session (a hosted turn builds its own). The cloud session test no longer reads ANTON_VERIFIER_JEV from the shell. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Tests improvments
…752) * feat(session): ask at a per-turn time limit and hand back with a summary (ENG-4006) A host can set max_turn_seconds (0, the default, is none: desktop has no time limit). Past it the turn asks "Keep going / Stop here" like the step and token limits; Keep going adds one more window, not four, and every other outcome ends the turn with the usual summary, recorded as time_limit. Checked per round after the other two limits, and at the continuation gate, which a text-only continuation would otherwise skip. The question is asked outside the model's question budget. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(cloud_turn): take the turn's time limit from the request's limits (ENG-4006) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(session): stop asking at the time limit after three more windows (ENG-4006) The total is the limit plus three windows. Past it the turn hands back with a summary instead of asking again, so it never runs into the hosting side's ceiling or outlives its key. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): always ask at the time limit, whatever the message said (ENG-4006) The message's "don't stop" granted the time limit like the step and token limits, so no question was asked. On web the host extends the turn's key only when that question is answered, so the key died while the turn ran on, and a turn with nobody to ask (scheduled, channel) ran past the limit. The time limit now always asks; with nobody to ask it hands back. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): start no tool call once the time limit has passed (ENG-4006) The time limit was checked per round, and a round runs its tool calls one after another, so every cell in it could overrun the limit: three 15-minute cells put the question 45 minutes late, past the end of the web turn's key window, which is sized for one last cell. Each call now checks the limit before it starts; a call past it gets a "not run" result, like a call after an unanswered question, and the round's next check asks. After Keep going the model can send it again. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): count the time-limit hand-back's continuations per turn (ENG-4006) staging now counts continuations on the session (_turn_continuations) rather than in a local, so the per-round time-limit hand-back read a name that no longer exists and every time-limit test failed with NameError. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): ask at the time limit when a text answer follows calls it didn't run (ENG-4006) A cell that crossed the limit mid-round left the round's other calls not run, and the model could answer that in text, ending the tool loop before its per-round check came round again. The turn then ended `completed`, never asking and never saying the limit cut it short. After the tool loop, a turn that had calls refused for the time limit now asks: Stop here hands back with the time-limit summary, Keep going resumes the task. A turn that simply finishes past the limit, with nothing refused, is not asked. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): count only the last round's refused calls at the time limit (ENG-4006) The refused-calls flag behind the post-loop time-limit check was never cleared. A step-limit or spend-ceiling stop on a later round then asked the time question as well: the user was asked twice, the turn relabelled, the reply appended again, and Keep going there resumed a turn the user had just stopped. A turn that finished on its own after Keep going was asked again. The flag now covers only the round that ended the tool loop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): mark the time-limit question in its id (ENG-4006) The step limit, spend ceiling and time limit all asked under `limit:`, so the web host couldn't tell them apart and extended the turn's key, or answered Stop here when it couldn't, on every limit's Keep going. The time limit now asks under `limit:time:`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(session): a late Keep going gets a full window, and a later hand-back restores the answer (ENG-4006) Keep going past the time limit now adds one window from the answer, not from the old limit, so a long last cell and the wait don't eat it; the allowance still stops at four windows in all. A Keep going that resumes work after the answer the user read now marks a continuation as pending, so a later hand-back announces itself and the host puts that answer back. Short limits read in seconds or a singular minute. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let a running turn take input the user adds mid-task (ENG-4194)
ChatSession gains add_user_input / withdraw_user_input /
take_unread_user_inputs and an on_user_input_read callback. The tool
loop shows unread added text to the model at the next step boundary,
in the same user message as the tool results, so turns still alternate.
If the model finishes with added text unread, the turn takes up to two
more rounds instead of ending on an answer that ignores it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Take mid-task messages in hosted (pod) turns too (ENG-4194)
In hosted mode the agent loop runs in the scratchpad pod, so a message
the user adds while it works now reaches the turn the way an ask_user
answer does: an `addon` line on stdin hands it to
ChatSession.add_user_input, an `addon_withdraw` line takes it back. The
pod reports each outcome as an `addon` event (type response.addon) on
its protocol stdout, keyed by addon_id: added/refused, read when the
model sees it, withdrawn/already_read, and returned for anything still
unread when the turn ends, on every path and before the terminal event.
The stdin reader now runs for every turn, not only interactive ones.
Lines that land before the session is built or the turn starts are
refused rather than buffered, so the server keeps them as the next task.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Pin the stream shape of a message read mid-turn (ENG-4194)
A message read at a step boundary emits no continuation, each round's text
comes before its own tool call, and the read is reported after the tool
ran and before the next round's text. Hosts rely on that order to keep
narration out of the answer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Harden mid-turn user input after self-review (ENG-4194)
- Limits: a turn takes at most 20 added messages, 4000 characters each,
64 KB in all; past them add_user_input returns False and the pod reports
"refused", so cowork-server keeps the message as the next task. One
injected block carries at most 16000 characters; the rest follows at the
next step.
- Framing: each added message is wrapped in <user_addition id="marker">
with a marker drawn per turn, user_addition tags inside the text are
stripped, and the block stays a separate text block. The system prompt
says additions only arrive that way, never in tool output.
- Read timing: on_user_input_read (and the pod's "read" event) fires when
the model call that was shown the text completes. A failed call or a
cancelled turn leaves it unread; it is returned at turn end and taken
back out of the history.
- The completion check judges the request together with the added
messages, so a retraction counts.
- The extra round also stops on the live spend ceiling; a blank answer is
not appended before it; the "replaces your previous answer" clause is an
anton note the check's transcript leaves out.
- Pod: emits {"kind":"addon","type":"response.addon","state":"ready"} once
at turn start, before any model call, when the stdin reader runs;
remembers refused ids; a withdraw after the message was returned reports
"returned".
- Leftover added text is cleared when a new turn starts.
- Tests: an Event replaces the stdin.tell() poll and sleep; no private flags
set where a public path exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(settings): raise the built-in turn limits to safety limits max_turn_tokens 1,250,000 -> 25,000,000, max_tool_rounds 25 -> 250, max_continuations 3 -> 10. The old values stopped real work: in 30 days of production no runaway turn appeared, and the largest legitimate one used 34.4M tokens, 171 tool rounds and 6 continuations. These values would have stopped 6 of 9,145 turns. Hosts that want tighter limits set their own, as cowork-server already does. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * test(e2e): pin the round and continuation limits the cap scenarios exercise These scenarios script a turn to the round cap (25 + grace) or past 3 continuations, and relied on those being the defaults. Each now sets ANTON_MAX_TOOL_ROUNDS / ANTON_MAX_CONTINUATIONS to the value it scripts, so they keep testing the cap rather than the default. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * fix(session): no limit grant on a turn that can't ask anyone A turn whose message says "keep going" or "don't stop" got 4x its limit without asking, even when no person could be asked: a non-interactive cloud turn, or turn(). At the new defaults that is 100M more tokens or 1,000 more rounds on an unattended run. The message-based grant, like the question, now needs a host that can ask a choice question (an emitter and an elicitor that takes one); otherwise the turn hands back at the configured limit. Also pins the built-in defaults in a test, and drops "25-round" from the spend-ceiling notice's docstring. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * docs(settings): the step limit counts across the whole turn Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs(session): say where a turn with no one to ask grants nothing Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…rsion Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Keep MCP connections' tokens out of the Scratchpad env An MCP connection's access mode (none/read/write) only limits which MCP tools anton registers. Its token was still injected as a DS_* env var, so code in the Scratchpad could call the provider's API directly and get past that limit, e.g. a shared Notion page asking the agent to edit pages on a read-only connection. env_for (desktop and cloud vaults) and the published-artifact secrets now skip the secret fields of a connection whose _method is "mcp". Other fields, and every non-MCP connection, are unchanged. The MCP tools read the token from the vault, not the env, so they keep working. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 2e24061) * Keep the prompt and output scrubbing right for withheld MCP tokens env_for now leaves an MCP connection's secrets out of the Scratchpad, so: - the Connected Data Sources list no longer names a token variable the pad does not have, and the prompt examples use a non-MCP connection; - the turn's scrub map still holds the withheld values, so an MCP error that echoes the token is redacted as before. They go to the scrub map only, never into an env. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit e243372) * Update the system-prompt snapshot for the non-MCP credential example The only change in the assembled prompt is the example env var, which now names a non-MCP connection since MCP tokens are no longer injected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Publish an app's secrets by env_for's rule, and name the ones it doesn't get Publishing kept its own copy of env_for's filter, which still sent `_` bookkeeping fields (`_method`, `_access_mode`, ...) that a local launch drops; it now resolves each datasource through env_for, so launch and publish agree. An MCP connection's token is never given to an app, so a backend that reads it used to fail with a bare 401: publish now returns the withheld names (`withheld_secrets`), the publish message lists them, and the backend launcher logs them. A test pins that a withheld token's scrub marker never unscrubs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…549) * Add Notion MCP server URL and unrestricted tool access Notion bounds every MCP tool by the signed-in user's own permissions, so it needs no read/write table. An opt-in UNRESTRICTED_ENGINES set keeps the fail-closed default for every other engine. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Pass _method through the turn-key vault so cloud MCP wiring sees Notion TurnKeyDataVault kept only an allow-list of token-response fields, so auth's _method marker was dropped and discover_mcp_tools found no MCP connection on cloud. Like LocalDataVault, _-prefixed fields are kept out of the injected env vars. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * General fix for MCP tool calls on smaller models Two failures seen on MindsHub Air with any MCP connector (HubSpot, Notion): - It fills optional fields with "" or null, and servers reject them (Notion's cursor="", HubSpot's query.cursor=""). Empty fields are now left out at any depth unless the tool's schema marks them required. - It sometimes announces the connector call in do-nothing scratchpad cells instead of making it, until the spend ceiling (~1.2M tokens). After two such rounds in a row the next round is sent without the scratchpad tool, plus a note to call the tool it needs or answer. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Make Notion read-only by default, like HubSpot All 49 Notion tools came to ~57k tokens of schema resent on every LLM call, enough to stall MindsHub Air when HubSpot was connected too. Notion now lists its 13 read tools in TOOL_ACCESS (~5k tokens); writes need a write-mode connection. The unrestricted-engine exception is removed. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Drop empty optional MCP arguments for read tools only On a write tool an empty string or null can be a deliberate "clear this field", so it now reaches the server unchanged. The cursor="" problem the cleanup fixes is on read tools (searches and listings). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Fit Notion changes to staging's tool-output and note-label rules Classify Notion's read tools in TOOL_OUTPUT, which staging now requires for every read-tier tool. Label the idle-scratchpad notice with ANTON_NOTE instead of a SYSTEM: prefix, which staging now forbids. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Keep MCP connections' tokens out of the Scratchpad env An MCP connection's access mode (none/read/write) only limits which MCP tools anton registers. Its token was still injected as a DS_* env var, so code in the Scratchpad could call the provider's API directly and get past that limit, e.g. a shared Notion page asking the agent to edit pages on a read-only connection. env_for (desktop and cloud vaults) and the published-artifact secrets now skip the secret fields of a connection whose _method is "mcp". Other fields, and every non-MCP connection, are unchanged. The MCP tools read the token from the vault, not the env, so they keep working. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Keep the prompt and output scrubbing right for withheld MCP tokens env_for now leaves an MCP connection's secrets out of the Scratchpad, so: - the Connected Data Sources list no longer names a token variable the pad does not have, and the prompt examples use a non-MCP connection; - the turn's scrub map still holds the withheld values, so an MCP error that echoes the token is redacted as before. They go to the scrub map only, never into an env. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Idle scratchpad guard: review fixes - Empty or non-text `code` is not idle: a list or number made ast.parse raise TypeError past the tool loop's handler, and an oversized cell that arrives empty is not a stall. RecursionError/MemoryError count as real work, like a SyntaxError. - `progress('...')` of literals counts as idle, like `print`. - The guard only runs on turns with MCP tools; without them the scratchpad is the only way to the data. - The verifier's continuation call rebuilds the tool list, so a withheld scratchpad is gone for one round only. - The action is read with scratchpad_action, the definition dispatch uses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * List notion-query-data-sources as a Notion read tool It is the only tool that reads a single database's rows on every Notion plan; notion-query-multiple-data-sources needs 2+ data sources and a Business or Enterprise plan with Notion AI. Leaving it out meant a read-only connection could not query one database, while a write-mode connection registered it anyway. Its schema is now ~28.5k characters (it was 83k when it was left out on size), so the read set is ~13k tokens. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Update the system-prompt snapshot for the non-MCP credential example The only change in the assembled prompt is the example env var, which now names a non-MCP connection since MCP tokens are no longer injected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Recognise a rejected token from an MCP server The MCP SDK turns any HTTP 4xx into a JSON-RPC error with no status on it, so a 401/403 from the server was classified as transient: an expired or revoked token was retried and surfaced as a generic failure, never as "reconnect". McpSession now records 401/403 responses off the wire and raises McpTokenRejected (a McpPermanentError carrying status_code) when opening the session, listing tools or calling a tool fails after one. Other failures are classified as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…wn examples Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Accept an optional label and engine on each cloud datasource connection reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * List each cloud datasource connection with its label and engine in the prompt. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Write the invisible characters in the datasource tests as escapes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Parse a datasource block recorded from cowork-server's producer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Keep a space in a connection label as a space instead of removing it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Keep both connection field checks in the contract, tell the agent to write SQL for each engine, and drop the prompt wording test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…arse the block the current producer sends. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t-discover-or-access-followup [ENG-4204] Say anton alone cleans connection labels, and test the current producer's block
…rs-in-org-mode-store-definitions-per [ENG-3838] Let cloud turns build a custom connector template
The "This step can't be checked before it runs" card now shows the step's code, scrubbed, and why it was held: code the check can't read, a name that earlier unreadable code may have set, a restart after a possible startup file, a depth limit, or a failed check. Building the reasons is guarded, so an error there loses only the reasons, not the changes the check found. Co-authored-by: Minura Punchihewa <minurapunchihewa17@gmail.com>
ANTON_WRITE_APPROVAL_UNCHECKED picks which cells that run code the check can't read wait on a card. "all", the default, keeps today's behaviour. "own" holds only the cell that runs such code, not later cells that may use what it made or the first cell after a restart. "off" holds none of them. Cards for SQL, file and listed HTTP writes stay in every mode, and a cell the check itself fails on still waits.
…cked setting (ENG-4246)
A user message can carry text next to its tool_results, such as the self-check notes active ACC mode adds. Only assistant rows keep text.
Keep the text beside tool calls in saved history rows
…G-4227) (#790) * Add ArtifactWorkspace: the folder the artifact editor reads and writes through (ENG-4207) anton/core/artifact_editor/workspace.py replaces the generic yolo workspace with one that knows what an artifact folder holds: - map/refuse_reason hide and refuse housekeeping, generation inputs, bytecode, binaries, dependency/build dirs and .revisions/ (names from artifacts/internal_files.py). - Per-kind rules: html-app stays one page (no sibling files the preview would never serve); fullstack apps keep backend.py, requirements.txt, state_manifest.json (stateful) and static/. - read() masks the observed-data binding as one placeholder line and write() restores it byte-exact, refusing a removed, duplicated or foreign placeholder. Compiled presentation pages show only their program; edits outside it are refused, the program is validated with component_page.document, and presentation-program.js is kept in step. - Read cap on the masked size; all-or-nothing preload (40 files, 48 KB, 32 KB per file, code first); capped search without SIGALRM that runs patterns that could backtrack badly as literal text. - write() validates every path first, then stages and renames each file no-follow, preserving modes and rolling back on failure; snapshot() restores bytes, modes and absence. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the artifact editor's patch engine: read a model's patch, place it by content (ENG-4207) A fresh Python port of yolocoder's patch parser and placer, cross-checked against Mochi's newer TypeScript port, as anton/core/artifact_editor/patch.py. It reads the compact format, apply_patch (V4A) and unified diffs, ignores every line number, and places each edit by matching its text: exact, then trimmed, then HTML entities decoded, then whitespace runs squeezed. Ordered placement, idempotent resends, and the four mechanical repairs (cut-off line, bunched edits, fragment of a long line, stray removed lines) come across with the traced reasons for each. Every unplaced edit in the patch is diagnosed in one pass, with the file's own lines to copy, and files whose edits all placed are returned for writing even when another file failed. Line endings (CRLF) and the trailing newline of each file are kept. One deliberate difference from both ports: a block that matches in more than one place is never placed at its first match, however long it is; when ordered placement cannot resolve it, the model gets the candidates. The old anton/core/yolo/patch.py is left in place for now. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the artifact editor and the edit_artifact tool (ENG-4207) Changing an artifact that already exists ("make the chart blue", "the save button does nothing") now goes to one tool, edit_artifact, which makes the change in one patch-writing model conversation, checks the page in a browser, repairs it once if the check fails, and puts every file back if it cannot make the change safely. anton/core/artifact_editor/: * editor.py: ArtifactEditor, yolocoder's change loop as Mochi ported it, on anton's LLM client. read_files / search / apply_diff with their own budgets (8 / 5 / 4, 20 rounds), the small-app preload as synthetic read_files pairs, "unchanged since it was shown above", partial applies with the failed files as they now stand, task_complete cleared by any later call and cancelled by a failed one, the check only when the model says it is done, cut-off and empty-reply handling, the whole-file rewrite as the last resort, snapshot restore on give-up and on cancellation. The embedded data's shape is shown beside its placeholder, and two header slips real models make are read as meant. * checkers.py: one CheckResult for static checks (JSON, Python, inline scripts with node), the render-check service (page and URL mode), the desktop's Electron and the user's preview (ChatSessionConfig.preview_checker). Unverified is never passed. * prompts.py, records.py: the prompt and tools; a JSONL record that always keeps rejected patches and every run's outcome. anton/core/tools/edit_artifact.py: the tool, joined to the turn's page safety net (turn page copies, repair cells, out-of-repairs refusal, live-app relaunch, attribution). The artifact may be named instead of its slug. ANTON_ARTIFACT_EDITOR=off|on (default on) and ANTON_ARTIFACT_EDITOR_ROLE=planning|coding. With it on, the ARTIFACTS prompt sections, generate_artifact's description and the browser_check_failed repair instructions route changes and repairs to edit_artifact. Removes anton/core/yolo/ and its tests: nothing on staging used it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the artifact edit benchmark and its fixtures (ENG-4207) scripts/bench_artifact_edit.py runs 28 edit tasks (cosmetic, content, behaviour, seeded bug fix) on 7 artifacts generated with anton's own generate_artifact, in three arms with the same model: a real ChatSession turn with the editor off (today's path), the same turn with it on, and ArtifactEditor called directly. Each run is its own process with a throwaway HOME and workspace and a fresh copy of the artifact, and records time, model calls, tokens, the task's assertions, the check status, an independent render check of the result and files changed outside the task. The fixtures hold synthetic data only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Record the artifact edit benchmark results (ENG-4207) 28 edit tasks on 7 generated artifacts, three arms, two models, with a local render-check service for every arm. On mindshub_air the median edit turn went from 34.3 s (5 model calls) to 18.0 s (3), with the same assertion pass rate; on sonnet-5-5 from 31.6 s to 20.1 s. Per-run numbers and both the final-code and same-window tables are included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the quick-edit bar's router and direct edit (ENG-4227) A change typed in the artifact viewer's bar ("make the revenue line purple") no longer needs a whole agent turn. The server asks Jev whether the artifact's own files are enough; if they are, it runs the artifact editor directly, checks the page, and fixes it from the check's findings. Anything that needs new data, a query, an analysis or an answer goes to the agent as before. anton/core/artifact_editor/: * router.py: route_request(), one Jev choice question (edit | agent) on MindsHub's /v1/decisions, the wire format and 0.8 acceptance rule of yolocoder's edit router. Uncertain answers, timeouts, errors, no key, an empty artifact or message all route to the agent; it never raises. Optional name / data shape / earlier messages (describe_artifact()). * quick.py: run_quick_edit() on the #790 editor loop, with up to max_fix_rounds (2) failed checks sent back with their findings, earlier bar messages as context, progress stages, and a QuickEditResult (applied | escalate | failed | noop). quick_edit_llm_from_settings(). * editor.py / prompts.py: opt-in escalation=True offers hand_to_agent; the model uses it when the request needs data or analysis the files do not have, and every file is restored (status "escalated"). Opt-in max_fix_rounds, Outcome.fix_rounds, and Progress.stage() callbacks. edit_artifact's behaviour is unchanged. * llm/jev.py: decide() can retry once after MindsHub's 503 "policy temporarily unavailable", inside the same deadline (opt-in). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Quick edit only where a decisions model exists (ENG-4227) Most installs are not on MindsHub and have no decisions model to route with. For them the bar must never edit directly: every message goes to the normal agent turn, whose edit_artifact works with any provider. * settings: decisions_model (ANTON_DECISIONS_MODEL). Unset means "jev-1.13.0" when the LLM is MindsHub (minds-cloud, or a base URL on api.mindshub.ai), otherwise none; an explicit empty value turns it off. * decisions_available(settings) / decisions_config(settings) -> (base_url, api_key, model) | None, for hosts to gate on before routing. * route_request with no model answers agent / no_decisions_model without a request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the quick-edit bar benchmark and its results (ENG-4227) scripts/bench_quick_edit.py measures the bar live on MindsHub: * --routing: route_request on 72 labelled bar messages (tests/fixtures/artifact_edits/routing.json, 40 edit / 32 agent, on the #790 benchmark artifacts), with and without the artifact's name and data shape: accuracy, confusion matrix, latency. * --e2e: route + run_quick_edit + the #790 render-check checker on the #790 benchmark's 28 edit tasks, one process per run. * --escalation: the editor alone on the agent-labelled messages, to see whether it hands them back. Results (jev-1.13.0, mindshub_air): routing 142/144 with context, no agent request routed to the editor, median 0.65 s; the bar end to end median 14.2 s / p90 25.2 s (the #790 agent turn with edit_artifact: 18.0 / 27.9); 30 of 32 agent requests handed back, 1 noop, 1 applied. docs/benchmarks/artifact-quick-edit-2026-11.md has the details. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Offer quick edit only when the LLM provider itself is MindsHub (ENG-4227) decisions_config said "available" for someone on Anthropic or OpenAI with their own key: it used a MindsHub key left over in ~/.anton/.env and a MindsHub URL from minds_url. Those setups must send every bar message to the agent. Decisions are now available only when the planning provider is MindsHub: minds-cloud (its minds_url and minds_api_key, on api.mindshub.ai), or openai-compatible whose base URL host is exactly api.mindshub.ai (its openai_base_url and openai_api_key). An explicit ANTON_DECISIONS_MODEL picks the model but never borrows a URL or key from elsewhere; an explicit empty value still turns it off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Quick edit: route with what the page contains; check cloud turns in the user's preview (ENG-4227) Routing. Jev only saw an artifact's name, description, data shape and the user's last bar messages, so "make the unicorns poop rainbow" on a page whose quick edits had added flying unicorns went to the agent. * describe_artifact() adds an outline (about 1,500 chars) read from the page's code: file names, title, headings, visible labels, ids, classes, keyframes, function and constant names, and features (canvas, WebGL / three.js, charts, forms, animation, tabs...). A compiled component page is outlined from its presentation program, not the shared runtime. Embedded data and housekeeping files are left out. * route_request() takes outline, recent_changes (the last 6 revision summaries) and chat (the last 4 messages of the artifact's chat); history now carries both sides, the last 6 turns. The instructions say that outline and recent_changes show what the page contains and that a request about things on it, or plausibly part of its visuals, is an edit. Old arguments still work. * Routing set: 12 more messages about page content, on a new synthetic canvas fixture (night sky with unicorns) and the existing artifacts. Remote preview check. A cloud turn had no preview_checker, so every edit was unverified and the agent told the user browser rendering was not available. * cloud_turn/preview_check.py: with "preview_checks": true in the turn request, the turn emits {"kind": "artifact_check", "type": "response.artifact_check", check_id, artifact_slug, timeout_s} and waits for a {"kind": "check_result", check_id, body} stdin line. No answer, null, stdin closing or the turn ending reads as unverified. * edit_artifact: an unverified check is described to the agent as "could not be checked from here; the user's open viewer shows any errors", with no instruction to tell the user about browsers; it never says verified. An unverified edit of a page that was not failing records nothing, so no "may not display correctly" line is added under the reply. The editor's note for the user no longer mentions checks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Record the partial page-context routing run; reword the editor's note rule (ENG-4227) * docs/benchmarks: the routing re-run with page context stopped when MindsHub began answering HTTP 402 wallet_empty for the benchmark account. What completed is recorded (contract variant with the new instructions: 81/84, "make the unicorns poop rainbow" to edit at p 0.98); the full-context variant still needs a run. * prompts.py: the note-for-the-user rule is worded without "do not mention", which tests/test_note_labels.py bans (models read it as an injection). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Preload the edit_artifact tool in the cloud turn resident (ENG-4207) Every cloud turn imports it now that the tool is registered, so the resident loads it once before forking, as the preload test requires. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…ENG-4227) (#800) #790's last commit was made from a stale working tree and reverted the two commits before it. Their changes come back, unchanged from 4d53d33: - The router gets what the page contains (an outline of its code, its recent revisions, the chat and both sides of the bar's history), so "make the unicorns poop rainbow" on a page with unicorns is an edit. - Cloud turns ask the person's open preview whether an edited artifact runs (artifact_check out, check_result back on stdin, opt-in per turn with preview_checks). - An unchecked edit no longer adds "not verified in a browser" under the reply, and the tool result tells the agent not to bring checks up. - The routing fixtures, tests and the partial live routing run. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… a card budget (ENG-4233) (#796) * test(write-check): replay a labeled chat corpus and hold the check to a card budget * test(write-check): add the critic pass's chats and name freed unread cells and broken pairs * test(write-check): give the replay corpus a generic write API, exit 2 on a refusal, and fix two labels * test(write-check): count a startup hook a cell writes as a real write in the replay corpus * test(write-check): key the replay by cell id, list the budget exactly, and accept a reasoned rise
…-4328) Lets the benchmark maintainers rerun anton on Runta from this repo at a pinned commit: Harbor and Pier agent classes, the install and run scripts their workflow calls, and the patch their cost script needs to read anton's usage.
fix(write-approval): keep check stops and whole cards under the unchecked setting (ENG-4246)
A step that writes a .pth file or a sitecustomize/usercustomize module is held only as code the check can't read, so `off` let it run with no card. Every later Python process runs that hook unchecked, so approving the write is the only point anyone sees it. The check now marks such a step (`writes_startup_hook`) and no setting drops its card.
The auth catalog moved the haiku alias to claude-haiku-5-5, so every eval run stopped on the repoint check. The slot covers the Anthropic-native provider shape, which claude-haiku-5-5 still is. The gate test now reads the pin instead of repeating the old id.
Show why a step can't be checked, and let operators choose which of those cards to show
…fter a turn, and the browser as last resort (ENG-4138) (#785) * Make the browser the last resort after connectors (ENG-4138) The browser prompt and skill now put a connected connector or API first, offer to connect a service's connector instead of browsing, and keep the browser for no connector, a user's ask, a failed fetch, and what sites do better than APIs (tickets, travel, checkout). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: work in one tab, gather before filling a form, and list a task's closed tabs (ENG-4138) The browser now caps a task's tabs (max_tabs, default 2) and notes the tabs of a recycled session (parked_tabs). action=tabs says the limit and lists those tabs, and the prompt asks for one tab, with what a form needs gathered into notes first and filled in one pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: after a turn, Jev decides whether the task is done with its browser (ENG-4138) When a turn that used the browser ends, Jev reads the request, the answer and the open tabs, and either closes the browser (the router notes its tabs, which reopen on request) or keeps it and closes the side tabs. Nothing closes while the person holds or watches the browser, or when Jev is unsure (under 0.7) or unavailable. Bounded at six seconds, after the answer has gone out. Not on the desktop app's own browser. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: Jev picks the steps, with each element's containers and the open dialog first (ENG-4138) - Jev's step choice reads each element's containers ("27" in "Select dates › November 2026") and, with a dialog open, chooses only among its elements. - The shortlist matches numbers (a calendar's 27), the inputs' values as well as their names, and an ISO date also as its day, month and weekday. - browser_jev defaults to on: run takes Jev's confident steps (0.8) itself and hands anything else back to the model. Commit steps still wait for the user's Approve. - A reCAPTCHA or hCaptcha frame named on the page no longer hands the browser over: those are often invisible, and the host reports the visible ones as challenges. - The prompt says to open a site's search address directly when it takes one, and to give run every value it needs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: on a crowded page Jev picks the part of the page, then the element; shadow mode is gone (ENG-4138) - A page with more candidate elements than one choice carries (40), in several parts, gets two questions: the step's kind with which part of the page it acts in (by containers, each described by its size and a few element names), then which element within that part. A part or element Jev isn't sure of (under browser_jev_min_p) goes back to the model. - Shadow mode, which only recorded Jev's picks against the model's, is removed with its analytics event. A config that still says shadow reads as off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Analytics test: the browser's PostHog events without the retired shadow event (ENG-4138) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: Jev can keep what a page shows and go on, handing the kept pages to the model (ENG-4138) A new step, collect, for a page that shows what the goal needs when more is elsewhere: run keeps its text (up to 2,500 characters, eight pages a run) and asks again for the step that moves on, with collect left out for that page. The kept pages come back with the run's answer, marked as website data. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: review fixes for run, human checks and the tidy-up after a turn (ENG-4138) A human check mid-run hands back the kept pages, and kept page text is fenced. Jev types or selects an input only when it is sure which one, and never clicks a reCAPTCHA or hCaptcha widget. On a crowded page the element call rates the risk again and the higher rating gates the step. Keeping a page no longer counts against max_steps. An on-screen captcha frame the host didn't report is a human check again; an off-screen one is not. close() tidies the browser only after a turn that ended normally and used it, and a host that doesn't send viewers closes nothing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: trust the host's challenges, and match captcha frames by name only from a host that sends none (ENG-4138) Page.challenges is None when the host doesn't send the field. The router's pod always sends it and leaves out invisible checks such as reCAPTCHA v3's on-screen badge, so an on-screen reCAPTCHA or hCaptcha frame is matched by name only when challenges is None. The run loop still refuses a Jev step whose target is named like a captcha. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: tidy only after completed turns and never while watched, keep pages when a run step fails, and prompt fixes (ENG-4138) - The tidy-up after a turn runs only after a completed turn, closes nothing (side tabs included) while the person watches, and reads the session again before closing, skipping on any change in viewers or holder. - A step that fails mid-run hands back the kept pages and the steps taken ahead of the host's refusal. - run says which element is missing before it asks about the input. - The prompt: asking to open, go to or show a page opens it in the browser; a site's search address is the one allowed built URL; position clicks no longer say they wait for Approve. The browser skill's step 6 now says to work in one tab and go back by address. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * README: a click by position runs at once (ENG-4138) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: a host's own decision model picks the steps (ENG-4138) BrowserConfig.decisions takes a DecisionsTarget (the endpoint's full URL, its key and its model name) that a host builds in its own process. With it the browser asks that endpoint for its steps, for the router and the desktop app's browser alike, whatever the LLM and with Jev on MindsHub turned off; without it, Jev on the session's MindsHub connection as before. jev.decide takes a full `url` for it: such endpoints serve TypeSafe's decision body at paths of their own, so nothing is appended. Like the launch key, the target never travels in a turn request: from_dict never reads one and to_dict refuses a config that has one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Browser: a cloud turn asks its host's decision model through the host, and the desktop browser asks Jev (ENG-4138) A cloud pod must not hold the organization's decision model key, so BrowserConfig gains `decisions_proxy` (a DecisionsProxy: the full https URL of a route of the host's own and the model name, no key), which a turn request carries in its browser block. The tool posts each decision there with the session's MindsHub credential, the turn key the router takes too; the host checks it, maps it to the organization and forwards the decision with the key it keeps. Like the host's own target it wins over Jev on MindsHub and ignores the Jev opt-out. to_dict writes it and from_dict reads it back; a malformed one, or one with any other key, reads as none, and the browser still runs. `decisions` (with a key) still never travels. The desktop app's own browser now asks Jev on MindsHub when the LLM is MindsHub, as the router does: running locally is about needing no browser server, not about where its pages go. README and docstrings say so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(browser): a hand-over waits for what the person does and carries on by itself (ENG-4138) When the person had to act on the page (a policy box only they may tick, a sign-in), Anton ended its turn with "do it and tell me", and nothing happened until they typed. hand_to_user now takes wait_for: an element to be checked or filled, or the page to change. The host watches for it and hands the browser back by itself, as the person's Done, once they did it and left the page alone for a few seconds; Done still ends it any time. The tool keeps the turn open meanwhile and returns with the page and whether the element is checked or filled, never its value. A host that doesn't know wait_for gets the hand-over without it. Human checks wait for the page to change. The tool prompt and the browser skill say to hand over with wait_for instead of ending the turn, and never to affirm a policy or consent for the person. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(browser): hand-overs promise only what the host watches, and read pages honestly after (ENG-4138) * fix(browser): after a hand-over's Stop, don't promise that Done carries the task on (ENG-4138) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Lucas Koontz <lucas.emanuel.koontz@gmail.com>
Add a FrontierHarness Eval adapter under contrib/frontierharness (ENG-4328)
This branch was successfully deployed
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
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.
Release: staging → main
126 commit(s) queued for the next production release.
Changes
Contributors
Review checklist