PR #73008: fix(deepseek): expose V4 max thinking levels

• Repository: openclaw/openclaw
• Author: ai-hpc
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/73008

Description (problem / solution / changelog)

Summary

• Problem: Native deepseek/deepseek-v4-pro and deepseek/deepseek-v4-flash rejected /think max and /think xhigh before request construction.
• Why it matters: DeepSeek V4 stream handling already maps max reasoning requests, but users could not select those levels through OpenClaw thinking validation.
• What changed: Added a DeepSeek V4 resolveThinkingProfile with off|minimal|low|medium|high|xhigh|max, added regression coverage, and documented the supported levels.
• What did NOT change (scope boundary): No OpenRouter behavior, no core thinking resolver changes, no changelog entry.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #72893
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: The DeepSeek provider registered the V4 stream wrapper, but did not expose a provider resolveThinkingProfile, so thinking validation fell back to the base levels only.
• Missing detection / guardrail: Existing DeepSeek tests covered payload mapping after a level was accepted, but not provider-profile level exposure.
• Contributing context (if known): DeepSeek V4 request shaping already handled xhigh/max via the stream wrapper.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: extensions/deepseek/index.test.ts
• Scenario the test should lock in: DeepSeek V4 models expose xhigh and max; non-V4 DeepSeek models do not opt into the V4 profile.
• Why this is the smallest reliable guardrail: The bug was in provider profile registration, so the provider plugin test directly exercises the missing hook.
• Existing test that already covers this (if any): Existing tests covered V4 payload mapping, not level acceptance.
• If no new test is added, why not: N/A

User-visible / Behavior Changes

/think xhigh and /think max are accepted for native DeepSeek V4 Flash and Pro models.

Diagram (if applicable)

Before:
[/think max] -> [DeepSeek provider has no profile] -> [base levels only] -> [rejected]

After:
[/think max] -> [DeepSeek V4 profile] -> [max accepted] -> [stream wrapper maps request]

Additional live test result:

Human verified with Deepseek API Key:

DEEPSEEK_LIVE_TEST=1 OPENCLAW_LIVE_DEEPSEEK_MODEL=deepseek-v4-pro pnpm test:live extensions/deepseek/deepseek.live.test.ts -- --reporter=verbose

RUN  v4.1.5

✓ extensions/deepseek/deepseek.live.test.ts > deepseek plugin live > returns assistant text from the bundled V4 model catalog 1103ms
✓ extensions/deepseek/deepseek.live.test.ts > deepseek plugin live > accepts V4 thinking replay after a prior provider tool call 1299ms

Test Files  1 passed (1)
Tests       2 passed (2)
Duration    5.18s

Changed files

• docs/providers/deepseek.md (modified, +2/-0)
• docs/tools/thinking.md (modified, +1/-0)
• extensions/deepseek/index.test.ts (modified, +31/-0)
• extensions/deepseek/index.ts (modified, +17/-4)
• extensions/deepseek/models.ts (modified, +5/-0)
• extensions/deepseek/stream.ts (modified, +5/-5)