fix(#1361): preserve reasoning, tool calls, and partial output on Stop/Cancel

[bergeouss]

Problem

Hitting Stop during a streaming response loses both the AI's already-streamed reasoning and tool output — paid tokens gone with no way to recover.

Three distinct data-loss paths:

§A — Reasoning text lost on cancel
_reasoning_text is accumulated in a thread-local variable inside _run_agent_streaming. On cancel, cancel_stream() never sees it — it goes out of scope when the thread is interrupted.

§B — Tool calls lost on cancel
_live_tool_calls is also thread-local. cancel_stream() only captures STREAM_PARTIAL_TEXT — tool calls the user saw on screen disappear after refresh.

§C — Reasoning-only streams produce no partial message
When the entire output is reasoning (no visible tokens), the regex strip of thinking blocks returns empty string. The if _stripped: guard skips the append entirely — only *Task cancelled.* survives.

Fix

Mirrors the same pattern as the existing STREAM_PARTIAL_TEXT shared dict:

1. STREAM_REASONING_TEXT — new shared dict in api/config.py, populated in on_reasoning() and the reasoning branch of on_tool(), read in cancel_stream()

2. STREAM_LIVE_TOOL_CALLS — new shared dict, populated on tool.started and tool.completed, read in cancel_stream()

3. §C fix — cancel_stream() now appends a partial assistant message when _stripped is empty but reasoning or tool calls exist, with the reasoning as a reasoning field and tool_calls as a tool_calls field

Tests

8 new regression tests in tests/test_issue1361_cancel_data_loss.py:

• ✅ Cancel with reasoning-only → reasoning preserved
• ✅ Cancel with reasoning + partial tokens → both preserved
• ✅ Cancel without reasoning dict → works as before (no regression)
• ✅ Cancel with tool calls → tool_calls preserved
• ✅ Cancel with tools + text → both preserved
• ✅ Reasoning-only creates partial message (§C)
• ✅ Tools-only creates partial message (§C)
• ✅ No reasoning, no tools, no text → only cancel marker (no change)

Existing tests unaffected: #1298 (10/10) and #1195 (5/5) all pass.

Closes #1361

[nesquena-hermes]

Shipped in v0.50.251 (merge 031feda3, tag https://github.com/nesquena/hermes-webui/releases/tag/v0.50.251). Thanks @bergeouss — solid foundation work on a real correctness gap.

Live at https://github.com/nesquena/hermes-webui/releases/tag/v0.50.251.

What landed

Your fix for #1361 ships clean — three distinct data-loss paths on Stop/Cancel are now closed:

• §A Reasoning text accumulated in thread-local _reasoning_text is now mirrored to STREAM_REASONING_TEXT shared dict, visible to cancel_stream()
• §B Live tool calls similarly mirrored to STREAM_LIVE_TOOL_CALLS
• §C Reasoning-only streams (no visible tokens) now produce a partial assistant message, not just a *Task cancelled.* marker

The pattern conformance to STREAM_PARTIAL_TEXT from #893 was excellent — same dict-creation, same _live_config fallback, same cleanup in _periodic_checkpoint.

Pre-release fix (Opus pass-2 MUST-FIX)

One critical correctness bug Opus caught:

Your fix stored captured tool calls under the message key tool_calls. That key is whitelisted by _API_SAFE_MSG_KEYS in api/streaming.py so the sanitize-for-API path forwards them to the next-turn LLM call. But the captured entries use the WebUI internal shape — they don't have OpenAI/Anthropic's id + function: arguments envelope. Strict providers would 400 on the malformed entries.

Concrete failure scenario: user starts a turn, agent does some tool calls, user hits Stop, then types a follow-up. The next API call would 400 because the partial message's tool_calls don't match the API contract. So the very cancel-then-continue scenario your PR aimed to improve becomes a hard fail.

Fix at f53556b: renamed the key to _partial_tool_calls (underscore-prefixed private key NOT in _API_SAFE_MSG_KEYS, so sanitize correctly strips it). Updated static/messages.js to also recognize _partial_tool_calls for UI rendering. Added tests/test_pr1375_partial_tool_calls_sanitize.py with 2 sanitize-roundtrip tests pinning the invariant.

This is a "make the structural shape itself enforce the invariant" fix — better than a comment because the field name being non-whitelisted means any future refactor that moves the partial-message logic still gets the right behavior automatically.

Tests

8 of your contributor tests + 2 sanitize-roundtrip tests = 10 total pinning the cancel-stream behavior end-to-end.