feat(llm): retain + replay plain-text reasoning for chat families (GLM/Kimi/DeepSeek/Qwen/Nemotron) by furgalep · Pull Request #301 · NVIDIA-NeMo/labs-OO-Agents

furgalep · 2026-09-07T06:43:35Z

Summary

Generalizes reasoning retention beyond OpenAI Responses (#261/#268) to chat-family reasoning models — GLM, Kimi, DeepSeek, Qwen, Nemotron — via plain-text reasoning_content, opt-in per alias. Stacked on #268 (fix/codeact-append-only-264).

ModelConfig.retain_reasoning — per-alias registry switch, default off (retained reasoning costs context on every subsequent turn)
Capture — the actor stores reasoning_content on terminal-text LLMOutput; CodeAct attaches it (with provenance) to the first ToolCallEvent of tool turns
Replay — OpenAIProviderFormatter emits reasoning_content on historical assistant turns (tool-call and text branches), provenance-gated: never sent to a different model family
model_family() fix — strips gateway/litellm routing prefixes (openai/, azure/, nvidia/, ...) before matching, and recognizes glm/kimi/deepseek/qwen/nemotron

Security/regression note (important)

Gateway ids like openai/nvidia/zai-org/glm-5.3 previously matched "openai" in model and were tagged openai — so the #268 provenance gate would not have blocked OpenAI-encrypted reasoning from being sent to GLM/Kimi through the gateway. This PR fixes the tagging (verified: openai/nvidia/zai-org/glm-5.3 → glm, openai/nvidia/moonshotai/kimi-k2.6 → kimi, openai/azure/openai/gpt-5.6-sol → openai).

Validation

Live probe: GLM-5.3 surfaces reasoning_content and the gateway accepts it replayed on a historical assistant message (turn-2 answer correct). Kimi not tested live (NVIDIA_INTERNAL_API_KEY unavailable in this environment) — presumed same gateway semantics, flagged for review.
Tests: model_family gateway-prefix units (2), replay-gate units (3), capture on/off (2)
1405 passed across strategies/config/formatter/runtime/unifiedllm; Ruff lint + format clean

Stack

#261 (retain Responses state) → #268 (append-only + provenance gate) → this PR. Merge in order; retarget as each lands.

🤖🤖🤖

Summary by CodeRabbit

New Features
- Added an optional setting to retain and replay plain-text reasoning in supported chat-model conversations.
- Reasoning is replayed only when it comes from the same model family, helping prevent cross-model context mixing.
- Reasoning retention is disabled by default and does not affect OpenAI Responses models’ encrypted reasoning.
- Improved recognition of model families when models use gateway or provider prefixes, including GLM, Kimi, DeepSeek, Qwen, and Nemotron models.

render_resume_picker, _preview_lines, and _wrap had no production callers: the picker renders through the shared browser, and the transcript owns the preview viewport. They moved into the test package as resume_picker_snapshot.py (a deterministic fixture importing the production clip/fragment helpers), so the thirteen snapshot assertions keep running against the same model state without shipping dead rendering code. The superseded model scroll state went with it: ResumePickerModel.scroll_preview and every preview_offset write (the live picker scrolls the transcript directly). The model scroll test now exercises the real surface — scrolling moves the transcript top row and row changes are visible. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

Event detail rendering walks the same markdown up to three times per paint (match lines, occurrences, highlighted output) and again on layout measurement — six-plus Rich renders of the same payload per keystroke. _styled_detail_lines now caches per (row, width), keyed by id(row) but retaining the row alongside the entry so a recycled id can never alias a fresh row onto stale lines. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

`except Exception` cannot catch asyncio.CancelledError (a BaseException in Python 3.13). A /model probe cancelled after _begin_model_validation took over the startup-prompt ownership skipped _mark_model_check_failed, so deferred prompts were never rejected or released and queued silently forever. The ownership resolution now runs on cancellation too before re-raising. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

The arm/confirm docstring promised a gesture window that never existed: an f armed minutes earlier still fired on the next f. The pending arm now carries its timestamp and expires after ten seconds — a stale arm re-arms instead of firing. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

The test asserted cleanup had not yet started between the cancel request and the very next line — a scheduler interleaving that only holds when the machine is idle. Under load, cancellation can legitimately begin there and the test failed intermittently. The contract under test is the synchronous status acknowledgement, which the retained assertions prove; the mid-flight ordering check is gone. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

test_resume_list_fills_its_pane assigned the SessionManager classmethods with bare classmethod objects — no monkeypatch, no teardown — so its twelve fabricated sessions (s00000001...) leaked into every later test. The memory-sidecars test then failed intermittently in full-suite runs (it asserts list_sessions() returns only its own session). The fixture now patches through monkeypatch, restoring the classmethods on teardown. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>

ResumePicker.__init__ duplicated ~85 lines of ExplorerBrowser.__init__ nearly verbatim: the buffer and search chrome, option controls, header controls, and the build_fullscreen_browser call. The base now provides the extension points — _create_list_control and _option_window_width factory hooks plus shared search label/close controls — and the picker's construction collapses to its model, its view facade, and super().__init__ with two hook overrides. The picker's duplicate invalidate() is gone too: the base one covers every control either constructs. Signed-off-by: Paul Furgale <pfurgale@nvidia.com>