- formatBriefTimestamp (5): force LC_ALL=en-US + TZ=UTC for CI
- EFFORT_LEVELS (1): update to 5 levels (added xhigh)
- effort description (1): remove stale 'no model names' test
- prompt caching (1): fix input_tokens expectation to match
Anthropic convention (non-cached only)
- ExecuteTool: rename runner to .runner.test.ts so bun finds it
- localVault/store.ts: synced from upstream
Remaining 4 failures are test isolation issues (compound command)
and feature flag mismatch (CtxInspectTool) — all pass in isolation.
DeepSeek v4 requires reasoning_content to be echoed back even when
empty, otherwise the API returns HTTP 400. The implementation already
handles this correctly (see convertMessages.ts L253-257 comment),
but the test expected undefined instead of empty string.
The buildDiffableContent and buildPrevDiffableContent fields were closures
capturing full system prompt and tool schema arrays (~300KB each). With 10
map entries × 2 closures, this held ~6MB of GC-unreachable memory.
Since recordPromptState already serializes the same data for hashing,
pre-computing the diffable content string has negligible marginal cost.
OpenAI's prompt_tokens includes cached tokens, but Anthropic's
input_tokens semantic excludes them. The adapter was mapping
prompt_tokens → input_tokens verbatim, causing downstream code
(cache hit rate, cost, autocompact) to double-count.
When ANTHROPIC_BASE_URL points to a non-Anthropic endpoint (e.g.
DeepSeek), the JSON-formatted user_id containing {, ", : characters
fails validation against ^[a-zA-Z0-9_-]+$. Send only the hex device_id
for third-party providers.
* feat: Add DeepSeek thinking mode support for OpenAI compatibility layer
- Add DeepSeek reasoning models support (deepseek-reasoner and DeepSeek-V3.2)
- Automatic thinking mode detection based on model name
- Inject thinking parameters in request body (both official API and vLLM formats)
- Preserve reasoning_content in message conversion for tool call iterations
- Extract buildOpenAIRequestBody() for testability
- Treat multimodal inputs (e.g. images) as new turn boundaries
- Fix env var cleanup in tests to prevent state leak
Signed-off-by: guunergooner <tongchao0923@gmail.com>
* docs: update contributors
---------
Signed-off-by: guunergooner <tongchao0923@gmail.com>
Co-authored-by: guunergooner <18660867+guunergooner@users.noreply.github.com>
Fixes#168
OpenAI requires that an assistant message with tool_calls be immediately
followed by tool messages. Previously, convertInternalUserMessage
output user content before tool results, causing 400 errors.
Now tool messages are pushed first.
Add xAI Grok as a new API provider. Reuses OpenAI-compatible message/tool
converters and stream adapter with Grok-specific client and model mapping.
Default model mapping:
opus → grok-4.20-reasoning
sonnet → grok-3-mini-fast
haiku → grok-3-mini-fast
Users can customize mapping via:
- GROK_MODEL env var (override all)
- GROK_MODEL_MAP env var (JSON family map, e.g. {"opus":"grok-4"})
- GROK_DEFAULT_{FAMILY}_MODEL env vars
Activation: CLAUDE_CODE_USE_GROK=1 or modelType: "grok" in settings.json
Also integrates with /provider command for runtime switching.
Two fixes for OpenAI-compatible provider compatibility:
1. Sanitize JSON Schema `const` → `enum` in tool parameters.
Many OpenAI-compatible endpoints (Ollama, DeepSeek, vLLM, etc.)
do not support the `const` keyword in JSON Schema. Recursively
convert `const: value` to `enum: [value]` which is semantically
equivalent.
2. Force stop_reason to `tool_use` when tool_calls are present.
Some backends incorrectly return finish_reason "stop" even when
the response contains tool_calls. Without this fix, the query
loop treats the response as a normal end_turn and never executes
the requested tools.
