* fix(providers): auto-clamp max_tokens on model rejection + fix verify for reasoning models
When OpenAI-compat models reject max_tokens as too large (e.g. gpt-3.5-turbo
supports 4096 but we send 8192), parse the model's stated limit from the 400
error, clamp the value, and retry once. This fixes agent creation for models
with lower output token limits without hardcoding model names.
Also increase the provider verify endpoint's max_tokens from 1 to 50 so
reasoning models (gpt-5, o-series) have enough headroom for internal
reasoning during the check call.
Closes#248, closes#245
* refactor(providers): extract chat retry closure + fix clamp log key
- Extract duplicate retry closure into chatRequestFn() to follow DRY
- Fix slog logging wrong key: body["max_tokens"] was nil for reasoning
models that use max_completion_tokens — now uses clampedLimit() helper
- Remove unnecessary _ = resp in provider verify endpoint
---------
Co-authored-by: viettranx <viettranx@gmail.com>
- Lead agents: auto-resolve team workspace as default (relative paths)
- Dispatched members: team workspace as default via req.TeamWorkspace
- Direct-chat members: own workspace default, team workspace accessible
- Add dataDir field to Loop/LoopConfig for global workspace root
- System prompt shows team workspace absolute path for model guidance
- Remove orphan task detector (superseded by post-turn dispatch)
- Log warning on OpenAI tool call argument parse failures
GPT-5 and o-series models reject the legacy max_tokens parameter and
require max_completion_tokens instead. Additionally, gpt-5-mini,
gpt-5-nano, and o-series models only support the default temperature (1).
This patch adds model-aware branching in buildRequestBody() to:
- Send max_completion_tokens instead of max_tokens for gpt-5*/o1/o3/o4
- Skip temperature parameter for gpt-5-mini/gpt-5-nano/o-series
Co-authored-by: Marcelo Emmerich <marcelo@agenticsystems.de>
Media tools (create_image, create_video, create_audio, read_audio,
read_video, read_document) routed API calls based on provider name
pattern matching (e.g. strings.HasPrefix(name, "gemini")). This breaks
when users give custom names to DB providers — a Gemini provider named
"chatgpt-sap-het" would be misrouted to the OpenAI-compat endpoint,
causing 404 errors.
Fix: carry the DB provider_type through OpenAIProvider, resolve it via
typedProvider interface in ExecuteWithChain, and inject as _provider_type
param for callProvider routing. Name-based heuristic kept as fallback
for config-file providers that don't have a DB type.
Co-authored-by: Luvu182 <208665161+Luvu182@users.noreply.github.com>
- Update go.mod and Dockerfile to Go 1.26
- Apply `go fix ./...` stdlib modernizations across 170+ files
- Add `go fix` to post-implementation checklist in CLAUDE.md
- Fix go fix misapplied rewrite in loop_history.go
Split agent summoning into two sequential LLM calls:
- Step 1: Generate SOUL.md (personality)
- Step 2: Generate IDENTITY.md + USER_PREDEFINED.md using SOUL.md as context
This enables real-time progress updates in the UI as each file completes.
On retry, previously generated files are detected by comparing against
the default template and skipped, so only missing files are regenerated.
Also increase HTTP client timeout from 120s to 300s for both OpenAI and
Anthropic providers. Some models (e.g. Kimi) with deep reasoning can
take 160s+ for non-streaming requests, which exceeded the 120s limit
and caused cascading retry failures until the context deadline expired.
Default bufio.Scanner limit (64KB) can truncate large SSE lines from
tool call arguments or thinking content, causing silent parse failures.
Set buffer to 1MB to match the scale of extended thinking responses.
Also remove a duplicate scanner.Err() check that was accidentally left in.
When stream_options.include_usage is enabled, the OpenAI API sends the
final usage chunk with an empty choices array. The streaming parser
skipped chunks with no choices before checking for usage data, causing
token counts to always be 0 in traces.
Fix: extract usage data before the empty-choices skip.
Also fixes two related SSE parsing issues:
- SSE data prefix now handles both "data: " and "data:" (some providers
like Kimi omit the space after the colon, per SSE spec)
- Added scanner.Err() check to surface stream read errors (timeout,
connection reset) instead of silently returning partial results
Co-authored-by: Luvu182 <208665161+Luvu182@users.noreply.github.com>
- SSE: handle "data:" without space (Kimi and other providers)
- Add scanner.Err() check to detect stream read failures
- Echo reasoning_content for thinking models (Kimi, DeepSeek)
- Add Thinking field to Message struct for reasoning passback
- Add GOCLAW_OPENAI_BASE_URL env var override (parity with Anthropic)
Implement full thinking mode system: per-agent thinking_level config (off/low/medium/high)
stored in other_config JSONB, per-provider param injection (Anthropic budget_tokens, OpenAI
reasoning_effort, DashScope enable_thinking+thinking_budget), Anthropic extended thinking with
streaming parse and tool use block preservation via RawAssistantContent, thinking token tracking
in trace spans, and Web UI with agent config selector, chat thinking block rendering, and trace
thinking token display. No DB migration needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add DashScope (Qwen) native provider with tools+streaming fallback
- Add Bailian Coding provider with hardcoded model list (no /v1/models API)
- Parse reasoning_content in OpenAI-compat streaming/non-streaming responses
- Emit ChatEventThinking events in agent loop for thinking models
- Add vision support for DashScope (qwen3-vl)
- Fix provider form dialog not updating API base URL when switching types
- Update README provider count from 11+ to 13+
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add tool loop detection (toolloop.go): tracks repeated no-progress tool calls
using SHA256 hashing of args+results. Warning at 5 identical calls, force stop
at 10. Prevents Gemini models from burning tokens in infinite loops.
- Inject AVAILABILITY.md negative context when agent has no team/delegation
targets, so models don't waste iterations probing unavailable capabilities.
- Fix Dockerfile: create /app/.goclaw directory so Docker volume initializes
with correct goclaw:goclaw ownership instead of root:root.
- Update collapseToolCallsWithoutSig comments for clarity.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Multi-agent AI gateway with WebSocket RPC, HTTP API, and messaging channel integrations.
Go port of OpenClaw with multi-tenant PostgreSQL, per-user isolation, security hardening,
and production observability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
