The previous comment claimed DeepSeek's API aliases v4-flash/v4-pro to
deepseek-reasoner, justifying the auto coercion via the legacy
reasoner's known 400 behavior. That had the alias direction inverted:
per DeepSeek's pricing page, deepseek-chat and deepseek-reasoner are
deprecated aliases for v4-flash's non-thinking and thinking modes
respectively, not the other way around.
The coercion itself is empirically correct, though. Live preflight
against deepseek-v4-flash with tool_choice=required produces a
deterministic HTTP 400 ("provider rejected forced tool selection") in
275ms. The behavior is consistent across the DeepSeek user community
- multiple downstream projects (pydantic-ai #5193, claude-code-router
#1378, opencode #24190, others) confirm V4 models reject forced tool
selection despite DeepSeek's chat-completion docs listing required as
a supported value. Server reality disagrees with documentation.
This commit updates only the comments in openai.go and openai_test.go
to point at the empirical evidence and the community confirmation.
The coercion behavior is unchanged.
DeepSeek's API server-side aliases deepseek-v4-flash and
deepseek-v4-pro to deepseek-reasoner, which rejects forced
tool_choice with HTTP 400 ("deepseek-reasoner does not support
this tool_choice"). Pulse's classifier then surfaced this as
"Selected model does not support Patrol tools," misdirecting
diagnosis to the model rather than the request shape.
supportsForcedToolChoice now returns false for any DeepSeek
client, so every DeepSeek model falls back to tool_choice
"auto" regardless of how DeepSeek routes the requested ID.
The ai-runtime contract is updated to match: the
provider-transport boundary now coerces forced tool_choice for
every direct DeepSeek model ID, not only unknown ones.
Patrol verified end-to-end: 20 tool calls, 9 findings, prior
runtime failure auto-resolved.
- coerce direct DeepSeek forced tool choices to auto for compatible requests
- preserve DeepSeek reasoning content across tool-result turns
- normalize Patrol runtime failure detail before run-history serialization
Keep the buffered EOF compatibility path for OpenAI-compatible streams that omit [DONE] but provide a terminal finish_reason, while rejecting truncated tool-call streams before they can produce executable tool calls.
Refs #1411
Refs #1412
Pass keep_alive=30s on both Chat and ChatStream requests so Ollama unloads the model shortly after Patrol or one-shot analysis bursts instead of refreshing the default 5-minute residency window on every request.
Tests assert both streaming and non-streaming Ollama request bodies include the keep_alive field.
Fixes#1425
Share OpenAI ChatStream finalization across DONE and clean EOF so buffered final chunks still produce tool calls and done events for OpenAI-compatible providers.\n\nFixes #1411
When local LLM servers (LM Studio, llama.cpp) receive tool definitions
but the model doesn't support function calling, they output internal
control tokens like <|channel|>, <|im_start|>, etc. instead of proper
responses.
This change detects these control tokens during streaming and returns
a clear error message explaining that the model doesn't support function
calling and recommending compatible models (Llama 3.1+, Mistral, Qwen).
This is better than the previous approach of offering a "disable tools"
option, which would have crippled Pulse Assistant/Patrol functionality.
Users need to use compatible models for the AI features to work properly.
Related to #1154
Some local models (llama.cpp, LM Studio) output internal control tokens
like <|channel|>, <|constrain|>, <|message|> instead of using proper
function calling. These tokens leak into the UI creating a poor UX.
This adds sanitization to strip these control tokens from both streaming
and non-streaming responses before they reach the user.
- Discovery: classify transient errors (429, timeout, connection refused, etc.)
and return IsError:true so models stop retrying rate-limited calls
- Agentic loop: detect identical tool calls repeated >3 times and block with
LOOP_DETECTED error, forcing the model to try a different approach
- OpenAI provider: skip tool_choice for DeepSeek Reasoner which doesn't support it
- Read-only classifier: fix curl -I case sensitivity (uppercase flags lowered),
add iostat/vmstat/mpstat/sar/lxc-ls/lxc-info/nc -z to allowlist,
fix 2>&1 false positive in input redirect detection
Gemini requires that model messages with function calls must be
immediately followed by user messages with function responses.
When message pruning or errors leave orphaned function calls,
Gemini rejects the request.
Add sanitizeGeminiContents() to:
- Strip orphaned function calls (keeping text content)
- Remove orphaned function responses without preceding calls
- Log when sanitization occurs for debugging
Provider updates across all supported backends:
- Anthropic: Better tool call handling
- OpenAI: Improved response parsing
- Gemini: Enhanced compatibility
- Ollama: Local model support improvements
Includes test updates for OpenAI provider.
Major new AI capabilities for infrastructure monitoring:
Investigation System:
- Autonomous finding investigation with configurable autonomy levels
- Investigation orchestrator with rate limiting and guardrails
- Safety checks for read-only mode enforcement
- Chat-based investigation with approval workflows
Forecasting & Remediation:
- Trend forecasting for resource capacity planning
- Remediation engine for generating fix proposals
- Circuit breaker for AI operation protection
Unified Findings:
- Unified store bridging alerts and AI findings
- Correlation and root cause analysis
- Incident coordinator with metrics recording
New Frontend:
- AI Intelligence page with patrol controls
- Investigation drawer for finding details
- Unified findings panel with actions
Supporting Infrastructure:
- Learning store for user preference tracking
- Proxmox event ingestion and correlation
- Enhanced patrol with investigation triggers
- Restore 'mini' mode for StackedDiskBar.
- Restore layout fixes (fixed table layout, mobile columns) for Docker and Hosts tables.
- Remove 'Ask AI' and AI context selection features.
- Docker: Use compact 'Cube' icon for Podman pods to prevent name obstruction.
- Docker: Show concise image names (strip registry URL).
- Backend: Include pending fixes for AI providers.
- Updated AI providers and tests for context/tenant awareness
- Refactored tool executor for multi-tenant state handling
- Added new tests for Docker control and update tools
- Add comprehensive tests for internal/api/config_handlers.go (Phases 1-3)
- Improve test coverage for AI tools, chat service, and session management
- Enhance alert and notification tests (ResolvedAlert, Webhook)
- Add frontend unit tests for utils (searchHistory, tagColors, temperature, url)
- Add proximity client API tests
- Add ChatStream method to all providers (Anthropic, OpenAI, Gemini, Ollama)
for real-time streaming of AI responses with tool call support
- Add StreamingProvider interface with StreamEvent types for content,
thinking, tool_start, tool_end, done, and error events
- Add notable models feature that fetches model metadata from models.dev
to identify recent/recommended models (within last 3 months)
- Add Notable field to ModelInfo struct to flag "latest and greatest" models
- Add SupportsThinking method to check for extended reasoning capability
The streaming support enables real-time AI chat responses instead of
waiting for complete responses. The notable models feature helps users
identify which models are current and recommended.
Users providing base URLs like "https://openrouter.ai/api/v1" were
getting HTML error responses because the client used the URL directly
without appending "/chat/completions".
- Normalize baseURL in NewOpenAIClient to ensure it ends with /chat/completions
- Fix modelsEndpoint() to derive /models from the normalized baseURL
- Add tests for URL normalization with various endpoint formats
The DisableDockerUpdateActions setting was being saved to disk but not
updated in h.config, causing the UI toggle to appear to revert on page
refresh since the API returned the stale runtime value.
Related to #1023
This update integrates learned baselines into the heuristic analysis to detect abnormal behavior and records significant events (migrations, restarts, spikes) for correlation analysis. Also fixed syntax errors in Ollama integration tests.
- Add persistent volume mounts for Go/npm caches (faster rebuilds)
- Add shell config with helpful aliases and custom prompt
- Add comprehensive devcontainer documentation
- Add pre-commit hooks for Go formatting and linting
- Use go-version-file in CI workflows instead of hardcoded versions
- Simplify docker compose commands with --wait flag
- Add gitignore entries for devcontainer auth files
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds RequestTimeoutSeconds to AI config (default 300s / 5 min).
Users with low-power hardware running Ollama can increase this
value in Settings to prevent timeouts on slower inference.
The issue was a SolidJS reactivity problem in the Dashboard component.
When guestMetadata signal was accessed inside a For loop callback and
assigned to a plain variable, SolidJS lost reactive tracking.
Changed from:
const metadata = guestMetadata()[guestId] || ...
customUrl={metadata?.customUrl}
To:
const getMetadata = () => guestMetadata()[guestId] || ...
customUrl={getMetadata()?.customUrl}
This ensures SolidJS properly tracks the signal dependency when the
getter function is called directly in JSX props.
rcourtman