fix(deepseek): clamp 'minimal' reasoning effort to 'low' instead of dropping it

[Drexuxux]

What

minimal reasoning effort was silently dropped by the DeepSeek profile (matched no arm → reasoning_effort omitted → DeepSeek fell back to its high default). So asking for the least reasoning gave you the most. Added the missing minimal → low clamp (the floor), matching the existing xhigh/max → max ceiling clamp.

Fix

• plugins/model-providers/deepseek/__init__.py — add minimal → low arm.
• tests/plugins/model_providers/test_deepseek_profile.py — add test_minimal_clamps_to_low.

Tests

pytest tests/plugins/model_providers/test_deepseek_profile.py -q → 32 passed. Revert the fix and the 3 new minimal cases fail (effort omitted → default high); with the fix they assert low.

[tonydwb]

Looks good. No obvious issues found.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the focused regression fix. The premise holds on current main: minimal is a valid Hermes effort (hermes_constants.py:794-822), while DeepSeekProfile only maps xhigh/max/ultra and low/medium/high (plugins/model-providers/deepseek/__init__.py:75-80). The standard chat transport applies this profile hook’s returned top-level kwargs to the request (agent/transports/chat_completions.py:579-592).

The proposed minimal → low mapping and case/whitespace regression coverage are consistent with that contract. Current GitHub main has not changed either touched path since the PR base in a way that conflicts; GitHub reports the PR cleanly mergeable.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs are associated with #30818. #30832 removes the default extra_body.thinking injection implicated by the successful curl/custom-provider comparison, while #30883 changes effort normalization, #54556 changes only the auxiliary-model default, and #65954 handles the separate minimal effort mapping.

Related pull requests

• fix(deepseek): make extra_body.thinking opt-in to unblock V4 first call (#30818) #30832 best fix — (+453/-22) — n/a: The recorded best-fix candidate guards extra_body.thinking behind explicit reasoning_config.enabled, making the default V4 first request match the clean request shape that succeeds, and adds regression coverage for that path.
• fix(deepseek): map low/medium reasoning_effort to high per API spec #30883 [closed] related — (+9/-4) — n/a: This normalizes low and medium reasoning efforts to high but leaves the default thinking injection unchanged, so it does not address the reported first-message cause; although closed, it remains relevant as a contributor-identified duplicate of feat: add DeepSeek V4 reasoning support #28945.
• fix: change DeepSeek default aux model from deepseek-chat to deepseek-v4-flash #54556 related — (+5/-5) — n/a: This changes only default_aux_model from deepseek-chat to deepseek-v4-flash, not the failing primary-model request payload. Its contributor keep-open review supports the independent aux-model change but requires replacing model-string pins with behavior-contract assertions.
• fix(deepseek): clamp 'minimal' reasoning effort to 'low' instead of dropping it #65954 related — (+26/-3) — n/a: This maps explicitly requested minimal reasoning effort to DeepSeek's low tier, preventing fallback to the server's high default, but it does not alter the unconfigured first-request payload from [Bug]: DeepSeek provider returns HTTP 400 on first message with deepseek-v4-flash (v0.14.0) #30818. The maintainer-bot keep-open verdict supports retaining this focused, independently salvageable regression fix.

Duplicates

#30883 duplicates the reasoning-effort normalization in #28945; #54556 and #65954 do not duplicate #30832 because they change independent auxiliary-model and effort-mapping behavior.

Suggested consolidation

Author action: rebase #30832 onto main and split out, if necessary, the opt-in thinking guard plus focused first-request regression coverage; this preserves the recorded best-fix path without closing it. Keep #54556 open with a salvage path that applies the contributor-requested behavior-contract tests, keep #65954 open with its focused minimal-to-low regression coverage, and leave #30883 closed as a duplicate of #28945.

Cross-PR triage: Reviewed 4 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 29 kB of PR diffs, 14 kB of issue/PR text, 4 kB of discussion (8 comments), 2 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[teknium1]

Your diagnosis was exactly right — dropping minimal fell through to DeepSeek's server default (high), the opposite of the ask. This is now fixed on main via the centralized effort vocabulary: agent/reasoning_effort.py (PR #90350) declares DeepSeek V4's set as low/medium/high/max and the shared clamp_effort() degrades minimal → low (nearest weaker level) in both the deepseek profile and the opencode-go relay path. Your earlier-submitted fix predates that refactor and matches its behavior exactly — closing as implemented on main, with credit for the correct diagnosis and direction.