fix(chat): persist aborted partial history

[penso]

Summary

• persist partial streamed assistant output on chat.abort so interrupted runs keep the visible assistant text and reasoning in session history
• persist assistant tool-call frames alongside tool results so tool-heavy interrupted runs round-trip back into prompt history
• cache aborted partial assistant messages in the web UI session-history store and add Rust plus Playwright regressions for abort recovery

Fixes #186.

Validation

Completed

• cargo test -p moltis-chat abort_persists_partial_stream_and_followup_reuses_it -- --nocapture
• cargo test -p moltis-chat send_sync_persists_tool_call_assistant_frames_for_history_replay -- --nocapture
• biome check --write crates/web/src/assets/js/websocket.js crates/web/ui/e2e/specs/chat-abort.spec.js
• cd crates/web/ui && ./build.sh
• cargo +nightly-2025-11-30 fmt --all -- --check
• cargo +nightly-2025-11-30 build --bin moltis
• cd crates/web/ui && MOLTIS_BINARY=/Users/penso/.superset/worktrees/moltis/issue-186/target/debug/moltis npx playwright test e2e/specs/chat-abort.spec.js
• cargo +nightly-2025-11-30 clippy -p moltis-chat --all-features --tests -- -D warnings

Remaining

• cargo +nightly-2025-11-30 clippy -Z unstable-options --workspace --all-features --all-targets --timings -- -D warnings
  blocked locally because llama-cpp-sys-2 configures GGML_CUDA=ON and this host does not have a CUDA toolkit available

Manual QA

1. Start a long-running chat response that emits text and at least one tool call.
2. Stop it before completion.
3. Send continue.
4. Confirm the aborted partial answer is still visible and the follow-up turn includes that partial context.

[greptile-apps]

Greptile Summary

This PR implements abort-recovery persistence for chat sessions: streamed assistant text and reasoning are buffered in a new ActiveAssistantDraft per-session map so that when chat.abort fires, the partially visible output is saved to the session store and broadcast back to the client. It also adds an Assistant { tool_calls } frame before every ToolResult entry so that values_to_chat_messages can correctly reconstruct prompt history for tool-heavy turns. The web UI's handleChatAborted handler is extended to cache the partial message and re-render it in the DOM, and bumpSessionCount for tool-call-end events is updated from 1 → 2 to account for the new frame.

Key points:

• The abort flow correctly resolves the session key before calling abort_run_handle, ensuring active_runs_by_session is still populated at lookup time.
• Cancelled Tokio tasks are dropped at their current .await point, so the normal-completion cleanup code does not race with persist_partial_assistant_on_abort — the design is sound.
• In both run_with_tools and run_explicit_shell_command, if appending the assistant tool-call frame fails, the code continues and appends the ToolResult anyway, creating an orphaned record that will break history replay on the next turn.
• handleChatAborted now has enough branches to trigger a cognitive-complexity lint, which is suppressed rather than addressed through helper extraction.
• The hasNonWhitespaceContent(partial?.audio || "") branch in hasVisiblePartial is dead code — the server never includes audio in partial messages.

Confidence Score: 4/5

• Safe to merge with one logic gap addressed: orphaned tool results when the assistant frame persist fails.
• The core abort-persist mechanism is architecturally correct and well-tested by both Rust integration tests and a new Playwright regression. The one functional concern — persisting a ToolResult even when the preceding assistant frame failed to write — is an edge case (storage failure) but would silently corrupt history replay. The JS changes are isolated to the abort handler and tool-call-end counter. Style issues (complexity suppression, dead audio check) are minor.
• Pay close attention to the tool-call frame + result persistence paths in crates/chat/src/lib.rs (both run_with_tools and run_explicit_shell_command).

Important Files Changed

Filename	Overview
crates/chat/src/lib.rs	Core implementation of abort-persist feature: adds ActiveAssistantDraft buffer, persist_partial_assistant_on_abort, tool call frame persistence, and two new integration tests. Logic is sound but has a gap: orphaned tool results can be written if the assistant-frame append fails.
crates/web/src/assets/js/websocket.js	handleChatAborted expanded with partial-message recovery logic; bumpSessionCount changed from 1→2 for tool-call ends. Function complexity lint suppressed; audio check in hasVisiblePartial is dead code; function mixes several concerns that could be split.
crates/web/ui/e2e/specs/chat-abort.spec.js	New Playwright test validates that an aborted broadcast keeps the partial assistant output in the UI and in the session-history cache. Test coverage is solid for the happy path.
crates/chat/Cargo.toml	async-stream added as a dev-dependency only, keeping it out of the production binary. No issues.

Sequence Diagram

sequenceDiagram
    participant C as Client JS
    participant S as LiveChatService
    participant DB as SessionStore

    C->>S: chat.send { text }
    S->>S: Insert ActiveAssistantDraft into active_partial_assistant
    S->>DB: append user_message
    loop Streaming
        S-->>C: chat delta events
        S->>S: draft.append_text / set_reasoning
    end
    S->>DB: append assistant_tool_call_frame
    S->>DB: append tool_result

    C->>S: chat.abort { sessionKey }
    S->>S: resolve_session_key_for_run
    S->>S: abort_run_handle — AbortHandle.abort, task dropped at current await
    S->>S: persist_partial_assistant_on_abort
    alt Draft has visible content
        S->>DB: append partial_assistant_message
        S->>DB: count session
        S-->>C: broadcast state=aborted with partialMessage and messageIndex
    else No visible content
        S-->>C: broadcast state=aborted no partial
    end
    S->>S: Remove active_thinking_text, active_tool_calls, active_reply_medium

    C->>C: handleChatAborted
    C->>C: cacheSessionHistoryMessage
    C->>C: bumpSessionCount if new messageIndex
    C->>C: Update DOM with partial content

Loading

Comments Outside Diff (1)

1. crates/chat/src/lib.rs, line 5459-5467 (link)

   Orphaned tool result if frame persist fails

   If store_clone.append(&sk_persist, &assistant_tool_call_msg.to_value()) fails (the warn-and-continue path), the code immediately tries to persist tool_result_msg anyway. A ToolResult stored in the history without a preceding Assistant { tool_calls: [...] } frame is an orphan — values_to_chat_messages will skip or mis-classify it, meaning the tool execution is silently dropped from the prompt history on the next turn.

   The same pattern appears in the error branch of run_explicit_shell_command (~line 5346). Since the whole point of this PR is to make tool-call/result pairs round-trip correctly, it would be safer to skip the tool-result append when the frame append fails:

   let frame_ok = store_clone
       .append(&sk_persist, &assistant_tool_call_msg.to_value())
       .await
       .is_ok();
   if !frame_ok {
       warn!("failed to persist assistant tool call frame; skipping tool result to avoid orphan");
   } else if let Err(e) = store_clone
       .append(&sk_persist, &tool_result_msg.to_value())
       .await
   {
       warn!("failed to persist tool result: {e}");
   }

_{Last reviewed commit: 1846476}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1846476007

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[codspeed-hq]

Merging this PR will not alter performance

✅ 39 untouched benchmarks
⏩ 5 skipped benchmarks¹

---

_{Comparing issue-186 (4bb869a) with main (aff4b62)}

Footnotes

1. 5 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[codecov]

Codecov Report

❌ Patch coverage is 88.76755% with 72 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/chat/src/lib.rs	88.76%	72 Missing ⚠️

📢 Thoughts on this report? Let us know!

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bbea499657

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4602943de1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 01bb8a0a2e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".