feat: implement diff handling for tool calls and enhance error detail…

[pikann]

Summary

• Agent Edit/Write-style tool calls rendered as raw stringified JSON args instead of a readable diff. ToolFallback now renders a real line-by-line diff (reusing DiffLineRow from content-diff.tsx) whenever a tool-call part carries diff data, falling back to the existing JSON view otherwise.
• eventsToThreadMessages now extracts diff content from two distinct event shapes, since edits can come from either backend path:
  • ACP-backed agents (Claude Code / Codex / Gemini CLI): ACPToolCallEvents that report edits as {type: "diff", path, old_text, new_text} content blocks.
  • Native OpenHands SDK agents: the file_editor tool's str_replace/create/insert commands, which report before/after content as old_content/new_content on the FileEditorObservation that follows the ActionEvent — a completely different shape with no diff-typed content block. Mirrors the SDK's own FileEditorObservation._has_meaningful_diff so view calls, errors, and no-op edits correctly fall back to the plain args view instead of showing an empty/misleading diff.
• services/ai-agent: _wait_for_done_or_stop now returns the ConversationErrorEvent detail alongside (stopped, errored, shutdown) so run_conversation can persist it to agent_conversations.error_message on a terminal ERROR/STUCK status. Previously this detail was only logged, leaving error_message NULL for every conversation that failed this way.

Fixes #315

Test plan

• bun run test in apps/web — 323 tests passing (incl. new coverage for both the ACP diff-block path and the native file_editor path: str_replace, create, view, and error cases)
• pytest tests/test_executor.py in services/ai-agent — 29 tests passing (incl. new coverage for error_detail propagation on a terminal ERROR status)
• tsc --noEmit / biome check on the changed frontend files — clean
• Replayed the real event stream of a conversation that uses the native (non-ACP) agent path, pulled from a local dev DB — confirmed 22 of 42 file_editor tool calls (the actual edits) now produce a diff artifact, while view calls and no-ops correctly do not

🤖 Generated with Claude Code

[pullfrog]

✅ No new issues found.

Reviewed changes — two complementary improvements: frontend diff rendering for ACP and native file-editor tool calls, and backend error detail propagation so agent_conversations.error_message is populated on ERROR terminal status rather than only logged.

• Render file diffs in tool-fallback cards — extractDiffBlocks parses ACP {type:"diff"} content blocks (tolerating both snake_case and camelCase), and extractFileEditorDiff handles native OpenHands file_editor observations. Tool-fallback shows diffs via DiffLineRow (exported from content-diff.tsx) through computeLineDiff computed inside useMemo. Diffs replace raw args when present.
• Propagate ConversationErrorEvent detail — _wait_for_done_or_stop returns a 4th error_detail tuple element, consumed by run_conversation and passed to update_conversation_status(error_message=...). Previously detail was only logged, leaving error_message NULL for every conversation that failed with ERROR/STUCK.

  ^{｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) (Kimi K2 not used — the program covers this model; add its provider key to run your pick) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — test-only additions that validate the error_detail propagation added in the prior commit.

• Unit test for _wait_for_done_or_stop error detail — new test_wait_for_done_or_stop_returns_error_detail_on_terminal_error creates a fake conversation in ERROR status and asserts the extracted detail matches the ConversationErrorEvent content. Existing tests updated to destructure the new 4-tuple.
• E2E assertions for error_message propagation — test_agent_reply_is_persisted_through_real_sandbox and test_agent_tool_call_then_reply_persists_full_flow now assert error_message=None on pause. test_llm_failure_marks_conversation_failed asserts "fake upstream failure" in error_message, closing the gap where the prior commit added the feature but didn't validate it end-to-end.

  ^{｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) (Kimi K2 not used — the program covers this model; add its provider key to run your pick) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes — performance guard and UX improvement: collapse long unchanged diff runs into context markers and skip expensive diff computation for very large files.

• Add collapseUnchangedContext to collapse long unchanged runs — new pure function in diff-utils.ts collapses stretches of unchanged lines beyond a configurable context window (default 3) into a single {type: "collapsed", count} marker, matching git diff's hunk convention. Both tool-fallback.tsx and content-diff.tsx use it so a one-line edit to a large file renders only the relevant context instead of every unchanged line.
• Add MAX_DIFF_LINES_PER_SIDE = 2000 guard in tool-fallback.tsx — skips computeLineDiff entirely when either the old or new text exceeds 2000 lines, showing a "diff too large" i18n message. The O(n·m) DP matrix for a multi-thousand-line file would otherwise freeze the tab.
• Add DiffCollapsedRow component and i18n keys — renders the collapsed-lines indicator using the new contentDiff.collapsedLines key across all 9 locales, plus contentDiff.diffTooLarge for the too-large guard.
• Thorough collapseUnchangedContext unit tests — 6 test cases cover short untouched runs, long mid-file runs with context on both sides, leading-only context, trailing-only context, short gaps that merge without collapsing, and empty input.

  ^{｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) (Kimi K2 not used — the program covers this model; add its provider key to run your pick) ｜ 𝕏}