feat: add a collapse assistant steps chat preference - #29115

Include renderer-detected failures in the collapsed badge

When process_signal returns an ordinary {success: false} result, its renderer promotes that response to an error even though is_error and status remain non-error (Tool.tsx:1113-1126). This predicate therefore leaves the failed tool out of failedCount, so with collapsing enabled the failure is hidden behind a normal Completed N steps summary with no warning badge. Compute the count using the same soft-failure classification used by the tool renderers.

Useful? React with 👍 / 👎.

Leaving this for the PR owner to decide. The badge counts protocol-level failures (is_error or status === "error"), which is also what the raw tool rows report; the soft-failure promotion is per renderer and differs by tool (process_signal checks success === false, start_workspace and the renderer at Tool.tsx ~414 check result.error, the subagent renderer has its own state), so counting them here means extracting a shared classifier from ~6 renderer sites in Tool.tsx. That is a refactor of the tool renderers rather than a fold change; if wanted it is better done as a follow-up so both the rows and the badge use one rule.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Move working-block assertions out of story plays

The new working-block story suites use play functions as assertion-heavy behavior tests, starting with this rendered-state check and continuing through keyboard, callback, expansion, and pagination assertions in both working-block story files. Pixel already captures the arranged visual state, while valuable behavior checks belong in Vitest; leave these plays setup-only and move the interaction assertions to React Testing Library tests.

AGENTS.md reference: site/AGENTS.md:L74-L74

Useful? React with 👍 / 👎.

Leaving this for the PR owner to decide. These plays were written before #29016 reworded FE1 on main, and the same pattern still exists on main (for example TogglesSendShortcut in this file). The two settings stories were migrated in a70cc67; the working-block suites cover keyboard, expansion, pagination and reading-position behavior that partly depends on real layout (the scroll story cannot move to jsdom), so migrating them is a separate mechanical pass rather than a fix for this PR's behavior.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Avoid DOM-geometry assertions in the scroll test

This test globally replaces offsetHeight and later asserts an exact scrollTop, so it verifies a hand-authored jsdom geometry model rather than observable browser behavior and can pass even when real scroller layout or anchoring is broken. Extract and unit-test the prepend decision as pure logic, or cover the actual scroll preservation in a browser-level test instead of mocking layout measurements.

AGENTS.md reference: site/AGENTS.md:L75-L75

Useful? React with 👍 / 👎.

Partly agreed. The observable behavior is covered against the real scroller and layout by the PrependIntoExpandedBlockKeepsReadingPosition story, which now exercises both the in-block page and the final page that also prepends the prompt row (it was red on the a70cc code and is green on a83e850). This unit test only pins the decision that a live row handing off to its persisted step must not scroll, which has no equivalent browser scenario yet; the mocked height is there because jsdom has no layout, not to model geometry. Leaving the thread open for the PR owner: either keep it as the guard's unit test or replace it with a browser story for the handoff.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Move cold-load behavior assertions into Vitest

These checks assert the intermediate loading state inside a Storybook play function, while the later checks also verify the loading-to-folded transition after an artificial timer. Pixel only captures the settled final state, so keep only the wait needed to arrange that screenshot here and cover the transition behavior with React Testing Library and Vitest instead.

AGENTS.md reference: site/AGENTS.md:L7-L7

Useful? React with 👍 / 👎.

Same question as the working-block story thread above: these plays predate the FE1 rewording in #29016. Leaving it for the PR owner to decide whether the story suites get migrated in this PR or in a follow-up pass.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Retain the successful preference value when refetch fails

When the preference PUT succeeds but the invalidation refetch fails, updatePreferenceSettings discards the successful response and leaves the old cached preference in place, so this controlled switch snaps back to its previous value even though the server saved the change. Unlike the already-handled initial-load failure, this post-save failure misrepresents a completed mutation; seed the cache from the mutation response before invalidating or otherwise preserve the successful value.

AGENTS.md reference: site/AGENTS.md:L11-L11

Useful? React with 👍 / 👎.

This uses the shared updatePreferenceSettings mutation from #/api/queries/users (mutate, then invalidate preferenceSettingsKey), the same as the other preference switches on this page, so a failed refetch after a successful PUT behaves identically for all of them. Changing the shared mutation to seed the cache from the response is a page-wide change rather than part of this preference; leaving it for the PR owner.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Compensate for stable-key read-file prepends

When pagination prepends more read_file calls into the first merged row of an expanded partial block, getDisplayMessageKey deliberately preserves that row's key even though its content grows above the reader. This equality check therefore skips the height correction and lets the current content jump downward. The fresh case beyond the earlier scroll fix is that a genuine prepend can grow a stable-key merged row, so prepend detection also needs to account for changes within that row.

Useful? React with 👍 / 👎.

Agreed this is a real gap, and a narrow one: it needs a read_file group at the block's first row to be split by a page boundary. getDisplayMessageKey keeps the group key stable on purpose, so the key comparison here cannot see the growth. The fix would be to identify the block by its member message ids instead of row keys (a prepend is then "the first id got smaller and the previous first id is still a member", and the live handoff has no previous id), which also changes the prop, the unit test and the story. Leaving it for the PR owner to decide whether it lands here or as a follow-up; note the same top-level group already grows without compensation on main when it is a scroller item of its own.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Scope fallback tool IDs to their message rows

When persisted tool-call parts omit tool_call_id, parseMessageContent generates IDs only from the part index, so two consecutive single-tool assistant messages both expose tool-call-0. This block-wide map then overwrites the first tool, causing the disclosure to report one step instead of two and potentially discard one row's failure state; scope fallback identities to the message or row while retaining deduplication for genuine protocol IDs.

Useful? React with 👍 / 👎.

The tool-call-<index> fallback only exists for persisted parts without a tool_call_id; chatd always writes ids for tool calls and results, and the map is keyed by id precisely so a call and its result from different messages merge into one step. Scoping fallback ids to their message would only change behavior for malformed transcripts. Leaving it for the PR owner; I would not change it here.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.

Stop the working timer during interruption

When a turn with persisted steps enters interrupting, the interrupt live row stays outside the fold, but this condition still marks the latest historical block live and keeps its Working for ... timer ticking beside the stopping indicator. The fresh state not covered by the requires-action fix is interrupting: deriveLiveStatus explicitly says this phase must not claim the agent is still producing output while the worker drains, so complete or pause the disclosure timer during interruption.

Useful? React with 👍 / 👎.

Deliberate: the PR treats interrupting as still working because the worker is draining the turn and the steps are not finished, and the completed label then uses the persisted timestamps, so the tick during the (short) interruption does not leak into the final duration. If the preference is to freeze the timer as soon as the interrupt is requested, that is a one-line change to isTurnActive; leaving the call to the PR owner.

Posted by Xum, an AI coding agent, on behalf of @ibetitsmike.