fix(deepseek): force tool_choice=required for thinking models + e2e regression suite

[roomote]

Opened on behalf of Elliott de Launay. View the task or mention @roomote for follow-up asks.

Related GitHub Issue

Addresses PostHog error 019e08dc: 24 occurrences of ConsecutiveMistakeError from deepseek-v4-pro (20) and deepseek-v4-flash (4) thinking models.

See also: #73 — follow-up to apply this fix holistically across all providers.

Description

Root cause: DeepSeek V4 thinking models sometimes return a plain text response instead of calling a tool when tool_choice: "auto" is sent, causing MODEL_NO_TOOLS_USED errors that accumulate into ConsecutiveMistakeError after 3 consecutive failures.

Fix (src/api/providers/deepseek.ts): Override tool_choice from "auto" to "required" inside DeepSeekHandler.createMessage when isThinkingModel is true and tools are present. Since attempt_completion is always included as a tool, every turn must end with a tool call — "required" is semantically correct here. Scoped to thinking models only to minimise blast radius; #73 tracks the global fix.

E2E regression suite (apps/vscode-e2e/src/suite/providers/deepseek-v4.test.ts): Four tests covering deepseek-v4-flash and deepseek-v4-pro × reasoning on/off. Each test verifies:

• Correct model is requested
• max_completion_tokens is set
• thinking and reasoning_effort params match the reasoning toggle
• Task completes via the read_file → attempt_completion tool loop
• Exact marker value is returned (no hallucination)

Fixtures use toolCallId matching for turn-2 to avoid cross-test count contamination from aimock's global match counters (documented in AGENTS.md).

Test Procedure

Mock mode (no API key needed):

USE_MOCK=true TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:run
# Expected: 4 passing

Record mode (requires DeepSeek API key):

DEEPSEEK_API_KEY=<key> TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:record

Unit tests:

cd src && pnpm test -- deepseek.spec

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved GitHub Issue (see above).
• Scope: Changes are focused on the linked issue.
• Self-Review: Thorough self-review performed.
• Testing: New and updated tests added.
• Documentation Impact: No user-facing docs require updates; AGENTS.md updated with fixture guidance.
• Contribution Guidelines: Read and agreed.

Screenshots / Videos

N/A — backend provider change, no UI impact.

Documentation Updates

• No documentation updates required.

apps/vscode-e2e/AGENTS.md updated with multi-turn fixture and toolCallId matching guidance.

Additional Notes

If another provider shows similar no_tools_used flakiness, apply the same targeted override pattern and reference #73.

Get in Touch

[roomote]

1 issue outstanding. Action required. See task

• Tighten the DeepSeek fetch interceptor to match the real host or origin instead of url.includes("api.deepseek.com"); the current substring check trips CodeQL and can match arbitrary URLs.
• Treat DEEPSEEK_API_KEY as a live-provider credential in apps/vscode-e2e/src/runTest.ts; otherwise TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:ci still forces aimock replay instead of exercising the real DeepSeek API.
• CI is still pending after the review wait window: platform-unit-test (windows-latest).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[roomote]

This adds the DeepSeek-specific record-mode key check, but the useMock gate below still ignores DEEPSEEK_API_KEY. Running DEEPSEEK_API_KEY=... TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:ci still starts aimock and injects AIMOCK_URL, so the new suite replays fixtures instead of exercising the live provider this PR is meant to validate. The DeepSeek key needs to count as a real-provider credential for DeepSeek-targeted runs.