Summary

• Problem: DeepSeek V4 thinking-mode tool calls require prior assistant reasoning_content to be replayed with tool-call history; missing it can make the DeepSeek API reject follow-up requests with 400.
• Why it matters: DeepSeek's official thinking-mode tool-call guide requires carrying reasoning_content through subsequent requests, including later user turns: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
• What changed: Adds deepseek-v4-pro and deepseek-v4-flash, defaults onboarding to deepseek/deepseek-v4-flash, enables DeepSeek thinking payloads, and backfills reasoning_content into replayed OpenAI-compatible chat messages.
• What did NOT change (scope boundary): Does not change core replay policy beyond the DeepSeek provider hook; no live API credentials or provider auth storage changes.

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 #
• Related #
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: The generic OpenAI-compatible conversion did not preserve DeepSeek reasoning_content in replay payloads for assistant messages that include tool_calls.
• Missing detection / guardrail: Existing DeepSeek tests only covered catalog/auth metadata and did not exercise realistic tool-call replay across multiple user turns.
• Contributing context (if known): DeepSeek's API contract differs from providers where reasoning can be omitted on replay; the official guide explicitly requires appending/replaying assistant messages with content, reasoning_content, and tool_calls for thinking-mode tool calls.

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: OpenAI-compatible request payload generation for DeepSeek multi-turn tool-call history preserves reasoning_content on every replayed assistant message, including tool-call turns and prior turns.
• Why this is the smallest reliable guardrail: It exercises the provider wrapper plus the real buildOpenAICompletionsParams conversion without requiring live DeepSeek credentials.
• Existing test that already covers this (if any): N/A
• If no new test is added, why not: N/A

User-visible / Behavior Changes

• Built-in DeepSeek catalog now exposes deepseek-v4-pro and deepseek-v4-flash with 1,048,576 context and 262,144 max output.
• DeepSeek onboarding now defaults to deepseek/deepseek-v4-flash.
• DeepSeek thinking-mode requests send thinking: { type: "enabled" } and reasoning_effort by default for reasoning-capable models.
• Replayed DeepSeek assistant messages preserve reasoning_content; empty reasoning strings are also replayed when present.

Diagram (if applicable)

Before:
DeepSeek tool call -> tool result -> follow-up request without reasoning_content -> possible 400

After:
DeepSeek tool call -> tool result -> follow-up request with reasoning_content + tool_calls -> accepted replay shape

Security Impact (required)

• New permissions/capabilities? (No)
• Secrets/tokens handling changed? (No)
• New/changed network calls? (No)
• Command/tool execution surface changed? (No)
• Data access scope changed? (No)
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: macOS
• Runtime/container: local OpenClaw checkout, Node via repo pnpm
• Model/provider: DeepSeek deepseek-v4-pro, deepseek-v4-flash
• Integration/channel (if any): Provider plugin / OpenAI-compatible chat completions
• Relevant config (redacted): N/A

Steps

1. Build a DeepSeek conversation history with user turn 1, assistant reasoning + tool call, tool result, assistant reasoning + second tool call, tool result, final assistant, user turn 2, and another assistant reasoning + tool call.
2. Convert that internal context through buildOpenAICompletionsParams.
3. Apply the DeepSeek provider stream wrapper.
4. Inspect the outgoing chat completions payload.

Expected

• Assistant messages in the outgoing payload include reasoning_content alongside content and tool_calls where applicable.
• Earlier-turn reasoning remains present in later-turn replay payloads.
• Empty reasoning_content is preserved as an explicit empty string when the stored thinking block exists.

Actual

• Covered by the new DeepSeek provider tests; payload assertions match the expected DeepSeek official tool-call replay shape.

Evidence

Attach at least one:

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Validation runs:

pnpm check:test-types
# passed

pnpm test extensions/deepseek
Test Files  1 passed (1)
Tests  13 passed (13)

Additional local check:

pnpm exec oxfmt --check --threads=1 docs/providers/deepseek.md extensions/deepseek/index.test.ts extensions/deepseek/index.ts extensions/deepseek/models.ts extensions/deepseek/onboard.ts extensions/deepseek/replay-history.ts extensions/deepseek/stream.ts
All matched files use the correct format.

Known unrelated local gate result:

pnpm check:changed
failed in typecheck extensions due existing extensions/codex AgentMessage type mismatch between root and extension-local @mariozechner/pi-agent-core installs.

Human Verification (required)

What you personally verified (not just CI), and how:

• Verified scenarios: DeepSeek catalog models, default model documentation/onboarding, reasoning payload injection, tool-call reasoning replay, multi-turn replay, and empty reasoning replay.
• Edge cases checked: Both deepseek-v4-pro and deepseek-v4-flash; assistant messages with tool_calls; prior-turn replay after a later user message; explicit empty reasoning string.
• What you did not verify: Live DeepSeek API call with real credentials; codex review --base origin/main was not run in this session.

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

If a bot review conversation is addressed by this PR, resolve that conversation yourself. Do not leave bot review conversation cleanup for maintainers.

Compatibility / Migration

• Backward compatible? (Yes)
• Config/env changes? (Yes)
• Migration needed? (No)
• If yes, exact upgrade steps: Existing configured DeepSeek model refs remain config-owned if explicitly set. New onboarding defaults use deepseek/deepseek-v4-flash.

Risks and Mitigations

• Risk: DeepSeek's official API behavior may evolve.
  • Mitigation: The replay behavior is provider-local and linked to the official thinking-mode tool-call contract.
• Risk: Generic OpenAI-compatible conversion shape changes could break the assistant replay index mapping.
  • Mitigation: Added a real conversion-chain test that asserts payload role order, tool-call IDs, and reasoning_content values.