Routing Reference

Astrolabe routes in four steps:

1. classify category and complexity
2. derive route modifiers
3. choose a lane
4. select candidates from the static manifest

Cheap low-risk turns can now skip both the model classifier and the verifier. Active OpenClaw work sessions can also reuse the last successful in-lane model when sticky execution metadata is present.

​

Lane defaults

Experimental branches exist behind an opt-in preview gate:

• research: grok420, gem31Pro
• vision: gem31Pro, gem31FlashLite

Those candidates only appear when both of these are true:

• ASTROLABE_ALLOW_PREVIEW_MODELS=true
• metadata.astrolabe.allow_preview=true

​

Key routing rules

• m27 is the default executor for most non-trivial core_loop, planning, orchestration, coding, research, and reflection requests.
• m25 only appears as a strict-budget or fallback path.
• Multimodal inputs promote the vision lane.
• Tool presence alone does not promote strict-json.
• Explicit structured output and schema-sensitive work promote the strict-json lane.
• Inside strict-json, m27 still gets the first attempt and specialists only activate after concrete validation failure.
• High-stakes or approval-sensitive work promotes the safe lane.
• Untrusted content plus tools cannot stay on cheap weak tiers.
• Sticky execution can keep the current workhorse stable across active OpenClaw sessions.

​

Profiles and guardrails

ASTROLABE_DEFAULT_PROFILE changes the final route after base lane selection:

• default: normal behavior
• strict-budget: allows m25 to replace m27 on cheaper low-risk routes
• safe-untrusted: pair with untrusted-content metadata for stricter safety floors
• low-latency: keeps the same lane logic but still routes onto m27
• max-capability: promotes more critical work upward

ASTROLABE_ROUTING_PROFILE still adjusts complexity before route resolution:

• budget
• balanced
• quality

Additional router controls:

• ASTROLABE_ALLOW_PREVIEW_MODELS
• ASTROLABE_STICKY_EXECUTOR_ENABLED
• ASTROLABE_SKIP_LOW_RISK_VERIFIER
• ASTROLABE_SKIP_LOW_RISK_CLASSIFIER

​

Escalation

Astrolabe only escalates non-stream responses. Important paths:

• m25 -> m27
• qwen35Flash/grok/qwenCoderNext/dsCoder -> m27
• m27 -> glm47Flash -> glm5 for strict JSON / schema recovery after validation failure
• m27 -> qwen35Plus or kimiThinking for research specialization
• m27 -> kimiK25 or qwen35Plus for multimodal specialization
• m27 -> sonnet -> opus only on critical or safe-lane escalation

​

Safety controls

• High-stakes detection can require explicit confirmation.
• Tool calls are inspected before Astrolabe returns them.
• Tool governance is capability-driven first and name-inferred only as a fallback.
• Read-only tool profiles can block mutating tool calls.
• external_comms, credential_access, remote_write, and financial_or_destructive actions require approval.
• untrusted_content blocks those dangerous capabilities outright and forces at least m27.
• Responses file and image URLs can be allowlisted.

​

Transparency

Use:

• GET /v1/models
• GET /v1/models?view=raw
• GET /v1/lanes

And inspect response headers:

• x-astrolabe-category
• x-astrolabe-lane
• x-astrolabe-initial-model
• x-astrolabe-final-model
• x-astrolabe-route-label
• x-astrolabe-sticky-applied
• x-astrolabe-sticky-reason
• x-astrolabe-verification-skipped
• x-astrolabe-transport-fallback-only