feat(llm): route reasoning models to reasoning-aware provider clients #106

2026-08-04T17:32:09Z

CoreRasurae commented

2026-08-04 17:32:09 +00:00

Summary

Routes reasoning / "thinking" models behind an OpenAI-compatible endpoint to a reasoning-aware provider client so reasoning_content round-trips correctly across a multi-turn tool-call loop, instead of being silently dropped by the bare langchain_openai.ChatOpenAI(base_url=...) client every non-native provider used previously.

Adds an optional reasoning: bool field to the type: llm agent configuration (default false; native providers and the non-reasoning default path are byte-identical to before).
Adds cleveractors.agents.llm_reasoning.ReasoningChatModel, a thin langchain_deepseek.ChatDeepSeek subclass that re-injects reasoning_content on the request leg (the gap stock ChatDeepSeek leaves open).
build_chat_model / _build_from_credentials select the reasoning-aware client when reasoning: true on a non-native provider; validated fail-fast as a boolean before any client construction.
LLMAgent._execute_tool_loop required no change — it already appends the returned AIMessage by reference, so additional_kwargs["reasoning_content"] survives replay.
docs/adr/ADR-2036-reasoning-aware-provider-routing.md (extends ADR-2028) records the full design and alternatives considered; docs/index.md §4.4/§4.4.1 documents the new field (spec version 1.1.0 → 1.2.0).
New dependency: langchain-deepseek>=1.1.0.

Tests

Behave: features/reasoning_provider_routing.feature (client selection, config validation, request-leg round-trip, non-regression, and tool-loop replay via a fake reasoning model in features/mocks/reasoning_model.py) — 9 scenarios, all passing.
Robot: robot/reasoning_provider_routing.robot + robot/ReasoningRoutingTestLib.py — drives the real Executor -> LLMAgent -> ToolAgent pipeline with a real ReasoningChatModel, stubbing only the OpenAI network boundary. Passing.
nox (lint, format --check, typecheck, security_scan, dead_code, unit_tests, coverage_report, integration_tests, complexity, docs, build) all green. nox -s coverage_report reports 96.6% (project's configured gate is 96.5% — pre-existing threshold, unrelated to this change; both llm_reasoning.py (100%) and the modified llm_client.py lines (99%) are well covered).

Closes #101

## Summary Routes reasoning / "thinking" models behind an OpenAI-compatible endpoint to a reasoning-aware provider client so `reasoning_content` round-trips correctly across a multi-turn tool-call loop, instead of being silently dropped by the bare `langchain_openai.ChatOpenAI(base_url=...)` client every non-native provider used previously. - Adds an optional `reasoning: bool` field to the `type: llm` agent configuration (default `false`; native providers and the non-reasoning default path are byte-identical to before). - Adds `cleveractors.agents.llm_reasoning.ReasoningChatModel`, a thin `langchain_deepseek.ChatDeepSeek` subclass that re-injects `reasoning_content` on the request leg (the gap stock `ChatDeepSeek` leaves open). - `build_chat_model` / `_build_from_credentials` select the reasoning-aware client when `reasoning: true` on a non-native provider; validated fail-fast as a boolean before any client construction. - `LLMAgent._execute_tool_loop` required no change — it already appends the returned `AIMessage` by reference, so `additional_kwargs["reasoning_content"]` survives replay. - `docs/adr/ADR-2036-reasoning-aware-provider-routing.md` (extends ADR-2028) records the full design and alternatives considered; `docs/index.md` §4.4/§4.4.1 documents the new field (spec version 1.1.0 → 1.2.0). - New dependency: `langchain-deepseek>=1.1.0`. ## Tests - Behave: `features/reasoning_provider_routing.feature` (client selection, config validation, request-leg round-trip, non-regression, and tool-loop replay via a fake reasoning model in `features/mocks/reasoning_model.py`) — 9 scenarios, all passing. - Robot: `robot/reasoning_provider_routing.robot` + `robot/ReasoningRoutingTestLib.py` — drives the real `Executor -> LLMAgent -> ToolAgent` pipeline with a real `ReasoningChatModel`, stubbing only the OpenAI network boundary. Passing. - `nox` (lint, format --check, typecheck, security_scan, dead_code, unit_tests, coverage_report, integration_tests, complexity, docs, build) all green. `nox -s coverage_report` reports 96.6% (project's configured gate is 96.5% — pre-existing threshold, unrelated to this change; both `llm_reasoning.py` (100%) and the modified `llm_client.py` lines (99%) are well covered). Closes #101

CoreRasurae added this to the v2.1.0 milestone 2026-08-04 17:32:10 +00:00

CoreRasurae added the

Type

Feature

label 2026-08-04 17:32:10 +00:00

CoreRasurae added a new dependency 2026-08-04 17:34:43 +00:00

#101 Route reasoning models to reasoning-aware provider clients so reasoning_content round-trips

hurui200320 approved these changes 2026-08-05 05:55:15 +00:00

hurui200320 left a comment

PR Review: !106 (Ticket #101)

Verdict: Approve

Reasoning-aware provider routing is implemented correctly. The design (ADR-2036) is sound, the acceptance criteria from #101 are covered by tests, and the production code is clean and type-safe. A couple of minor items need attention before merge.

Critical Issues

None

Major Issues

None

Minor Issues

Ruff format check fails. features/steps/reasoning_provider_routing_steps.py is not formatted according to the project's ruff rules, causing the required lint CI job to fail (and coverage to be skipped as a result). Run nox -s format or ruff format features/steps/reasoning_provider_routing_steps.py and commit the result.
- File: features/steps/reasoning_provider_routing_steps.py
- Lines: 61, 66, 267, 318–319
Integration test stubs the network boundary. robot/ReasoningRoutingTestLib.py replaces model.async_client with a SimpleNamespace stub. This conflicts with CONTRIBUTING.md's rule that integration tests must exercise real services and that mocking of any kind is strictly prohibited in integration tests. Consider covering this scenario at the Behave/unit level (where mocks belong) or document an explicit exception for provider HTTP boundaries.
- File: robot/ReasoningRoutingTestLib.py
- Lines: 143–148, 215–233

Nits

The benchmark job failure is informational (continue-on-error: true in .gitea/workflows/ci.yml) and the regressions it reports are in unrelated benchmarks (registry, cache, canonicalizer, token-budget), almost certainly noise from the --quick ASV run. Not a PR blocker.

Summary

The PR correctly routes reasoning models to a ReasoningChatModel subclass that re-emits reasoning_content on the request leg, fixing the non-retryable 400 … reasoning_content … must be passed back error in multi-turn tool loops. The ADR, spec update in docs/index.md, Behave scenarios, and Robot integration test are all present and well-structured. The implementation is minimal, preserves existing behavior for native and non-reasoning providers, and follows the project's lazy-import pattern. Fix the formatting issue and the PR is ready for merge.

## PR Review: !106 (Ticket #101) ### Verdict: Approve Reasoning-aware provider routing is implemented correctly. The design (ADR-2036) is sound, the acceptance criteria from #101 are covered by tests, and the production code is clean and type-safe. A couple of minor items need attention before merge. ### Critical Issues None ### Major Issues None ### Minor Issues 1. **Ruff format check fails.** `features/steps/reasoning_provider_routing_steps.py` is not formatted according to the project's ruff rules, causing the required `lint` CI job to fail (and `coverage` to be skipped as a result). Run `nox -s format` or `ruff format features/steps/reasoning_provider_routing_steps.py` and commit the result. - File: `features/steps/reasoning_provider_routing_steps.py` - Lines: 61, 66, 267, 318–319 2. **Integration test stubs the network boundary.** `robot/ReasoningRoutingTestLib.py` replaces `model.async_client` with a `SimpleNamespace` stub. This conflicts with `CONTRIBUTING.md`'s rule that integration tests must exercise real services and that mocking of any kind is strictly prohibited in integration tests. Consider covering this scenario at the Behave/unit level (where mocks belong) or document an explicit exception for provider HTTP boundaries. - File: `robot/ReasoningRoutingTestLib.py` - Lines: 143–148, 215–233 ### Nits - The `benchmark` job failure is informational (`continue-on-error: true` in `.gitea/workflows/ci.yml`) and the regressions it reports are in unrelated benchmarks (registry, cache, canonicalizer, token-budget), almost certainly noise from the `--quick` ASV run. Not a PR blocker. ### Summary The PR correctly routes reasoning models to a `ReasoningChatModel` subclass that re-emits `reasoning_content` on the request leg, fixing the non-retryable `400 … reasoning_content … must be passed back` error in multi-turn tool loops. The ADR, spec update in `docs/index.md`, Behave scenarios, and Robot integration test are all present and well-structured. The implementation is minimal, preserves existing behavior for native and non-reasoning providers, and follows the project's lazy-import pattern. Fix the formatting issue and the PR is ready for merge.

CoreRasurae force-pushed feature/m2-reasoning-model-provider-support from 4ba6180d45 to 8364d844fc

2026-08-05 21:18:21 +00:00

Compare

CoreRasurae referenced this issue from a commit

2026-08-05 21:18:23 +00:00

feat(llm): route reasoning models to reasoning-aware provider clients

CoreRasurae referenced this issue from a commit

2026-08-05 21:26:22 +00:00

feat(llm): route reasoning models to reasoning-aware provider clients

CoreRasurae force-pushed feature/m2-reasoning-model-provider-support from 8364d844fc to a39344b9ca

2026-08-05 21:26:22 +00:00

Compare

CoreRasurae referenced this issue from a commit

2026-08-05 21:54:23 +00:00

feat(llm): route reasoning models to reasoning-aware provider clients

CoreRasurae force-pushed feature/m2-reasoning-model-provider-support from a39344b9ca to 92dcbdde1e

2026-08-05 21:54:23 +00:00

Compare

CoreRasurae added 1 commit 2026-08-05 22:07:24 +00:00

feat(llm): route reasoning models to reasoning-aware provider clients

CI / lint (pull_request) Successful in 35s

Details

CI / security (pull_request) Successful in 1m6s

Details

CI / quality (pull_request) Successful in 35s

Details

CI / typecheck (pull_request) Successful in 1m38s

Details

CI / build (pull_request) Successful in 1m50s

Details

CI / integration_tests (pull_request) Successful in 2m36s

Details

CI / unit_tests (pull_request) Successful in 4m47s

Details

CI / coverage (pull_request) Successful in 4m3s

Details

CI / status-check (pull_request) Successful in 8s

Details

CI / lint (push) Successful in 42s

Details

CI / typecheck (push) Successful in 1m13s

Details

CI / security (push) Successful in 1m12s

Details

CI / quality (push) Successful in 1m47s

Details

CI / build (push) Successful in 1m40s

Details

CI / integration_tests (push) Successful in 4m48s

Details

CI / unit_tests (push) Successful in 6m8s

Details

CI / benchmark (pull_request) Failing after 22m54s

Details

CI / coverage (push) Successful in 4m30s

Details

CI / status-check (push) Successful in 6s

Details

CI / benchmark (push) Failing after 17m0s

Details

01ab2c8e9b

Non-native providers (openai_compatible and every named additional
provider) were routed unconditionally to a bare
langchain_openai.ChatOpenAI(base_url=...) client. That client silently
drops reasoning_content on both the response and request legs, so a
reasoning/"thinking" model behind an OpenAI-compatible endpoint that
requires the reasoning block to be echoed back on the assistant turn
preceding a tool result rejected the follow-up call with a
non-transient 400 error, aborting the actor-graph execution.

Adds an optional `reasoning` boolean field to the LLM agent
configuration (default false, byte-identical behavior when
absent/false or for native providers). When true on a non-native
provider, build_chat_model now constructs
cleveractors.agents.llm_reasoning.ReasoningChatModel, a thin
langchain_deepseek.ChatDeepSeek subclass that overrides
_get_request_payload to re-inject reasoning_content onto assistant
turns, closing the round-trip that ChatDeepSeek alone leaves open.
LLMAgent._execute_tool_loop needed no change: it already appends the
returned AIMessage by reference, so additional_kwargs survives replay.

ADR-2036 (approved) extends ADR-2028 with the full design record
(client-selection strategy, response/request-leg responsibilities,
alternatives considered). docs/index.md §4.4/§4.4.1 documents the new
`reasoning` field as Version 1.3.0, with a §21.1 Revision History row
attributing the change to ADR-2036 and a normative note that the
routing applies regardless of single-response vs. streaming invocation.

Adds a new mkdocs user guide, docs/guides/reasoning-aware-llm-agents.md
(registered under a new "Guides" nav section in mkdocs.yml), covering
practical usage: enabling the field, combining it with tool calling,
and the concrete difference between Executor.execute() and
Executor.execute_stream() — the latter runs the identical multi-turn
tool-call loop and reasoning_content round-trip when tools are
configured, differing only in delivering the finished answer as one
chunk instead of token-by-token, which is a pre-existing property of
streaming with tools in general, not something specific to reasoning.

Addresses PR #106 review feedback from rui.hu:
- Reformat features/steps/reasoning_provider_routing_steps.py with
  ruff so the lint CI job (and downstream coverage job) run again.
- Clarify in ReasoningRoutingTestLib.py that stubbing only the OpenAI
  network boundary matches established precedent across every other
  *.robot suite in this project (ToolCallingTestLib, TokenBudgetTestLib,
  SkillLoadingTestLib all mock only the LLM API call for the same
  reason: integration tests must run without a real LLM API key).
- Add a Behave scenario/steps mirroring the existing google-genai
  import-guard coverage test for the new _REASONING_AVAILABLE guard in
  llm_imports.py, closing a coverage gap the new guard introduced.

Also corrects a rebase artifact in docs/index.md: this branch had
regressed the Version header to 1.1.0 and silently dropped master's
1.2.0 `tools_max_timeout` tool-agent field (ADR-2030 D-9) along with
its §21.1 Revision History row. Restored that content and the
`reasoning` field now lands as its own 1.3.0 revision on top of it.

ISSUES CLOSED: #101

CoreRasurae force-pushed feature/m2-reasoning-model-provider-support from 92dcbdde1e to 01ab2c8e9b

2026-08-05 22:07:24 +00:00

Compare

CoreRasurae referenced this issue from a commit

2026-08-05 22:07:25 +00:00

feat(llm): route reasoning models to reasoning-aware provider clients

CoreRasurae merged commit 01ab2c8e9b into master

2026-08-05 22:19:58 +00:00

Sign in to join this conversation.