feat: ClawRouter v0.5.0 - Agentic routing, context-aware fallbacks, tool detection #10
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… UI" This reverts commit 6c32e50.
- Apply BlockRun design patterns: dark bg, subtle borders, JetBrains Mono - Add footer links: blockrun.ai, X/Twitter, GitHub repo - Change chart to benchmark comparison: ClawRouter vs Claude Opus 4 - Use uppercase labels with letter-spacing per BlockRun style
Users can now use short aliases instead of full model paths: - /model claude → anthropic/claude-sonnet-4 - /model kimi → moonshot/kimi-k2.5 - /model gpt → openai/gpt-4o - /model flash → google/gemini-2.5-flash Aliases: claude, sonnet, opus, haiku, gpt, gpt4, gpt5, mini, o3, deepseek, reasoner, kimi, gemini, flash, grok
When agenticMode is enabled, the router prefers models optimized for agentic workflows (Claude, Kimi K2.5) that continue autonomously instead of stopping and waiting for user input. - Added 'agentic' flag to model definitions - Marked Claude, Kimi K2.5, GPT-5.2, GPT-4o as agentic - Added agenticTiers config for agentic model preferences - Router uses agentic tiers when agenticMode: true - Added isAgenticModel() and getAgenticModels() helpers
When session persistence is enabled, ClawRouter maintains the same model selection across multiple requests within a session, preventing disruptive model switches during multi-step agentic tasks. - Added SessionStore class for tracking session -> model mappings - Sessions identified via X-Session-ID header - Configurable timeout (default: 30 minutes) - Auto-cleanup of expired sessions - Exported SessionStore, getSessionId, DEFAULT_SESSION_CONFIG
- Add 15th dimension: agenticTaskKeywords (67 multilingual keywords) - Auto-switch to agentic tiers when 2+ agentic signals detected - Supports file ops, execution, multi-step patterns, iterative work - No config required - works automatically Test results: - "what is 2+2" → gemini-flash (standard) - "build the project then run tests" → kimi-k2.5 (auto-agentic) - "fix the bug and make sure it works" → kimi-k2.5 (auto-agentic)
- Remove /dashboard endpoint from proxy - Remove generateDashboardHtml function - Keep getStats() and formatStatsAscii() for terminal display - Simpler architecture, no remote access issues
- Remove "Hello" and "Define photosynthesis" from SIMPLE tier tests (now route to MEDIUM with adjusted scoring weights) - Update REASONING tier test to expect deepseek-reasoner instead of o3 - Add local /v1/models endpoint that returns model list without upstream API call - Update fallback test to accept deepseek or gemini for SIMPLE tier - Bump to v0.4.9
- Add 9 new models: grok-4-fast-*, grok-code-fast-1, nvidia/gpt-oss-120b - Update tier config: REASONING uses grok-4-fast-reasoning ($0.20/$0.50) - Add nvidia/gpt-oss-120b as FREE fallback in SIMPLE tier - Empty wallet auto-fallback: use free model instead of throwing error - Update test expectations for new model routing
When model returns tool_calls instead of content (e.g., for function calling), ClawRouter was silently dropping the tool_calls data during SSE conversion. This caused OpenClaw to receive finish_reason: 'tool_calls' with empty content, resulting in no reply being sent to users. Now tool_calls are properly forwarded in the SSE stream.
- Remove 'View detailed dashboard' URL from /stats output (dashboard removed) - Skip content-encoding header when forwarding upstream response (fetch auto-decompresses but was forwarding gzip header with plain body)
When tools array is present in request, automatically switch to agentic tiers which use models better at tool use (claude-haiku-4.5, claude-sonnet-4, kimi-k2.5) instead of reasoning-focused models (deepseek-reasoner) that may generate malformed tool calls.
Add smart filtering that excludes models when estimated context exceeds their max context window. This prevents wasted API calls to models that will fail with context length errors. Changes: - Add getModelContextWindow() helper to models.ts - Add getFallbackChainFiltered() to selector.ts - Update proxy to filter fallback chain by context length - Update fallback tests for agentic tier fallback order
- Add tool detection section - Add context-length-aware routing section - Add session persistence section - Add /stats command documentation - Update tier model mapping (grok-code-fast, grok-4-fast-reasoning) - Add new xAI Grok 4 family and NVIDIA free models - Update roadmap with completed features
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Major update with intelligent routing improvements:
toolsarray is present, automatically switch to agentic tierstool_callsin streaming responses/statswith cost savings dashboardTest Results
Breaking Changes
None - backwards compatible.
Commits