Runtime-aware Skill routing for real Codex work
Classify Single, Phased, and Managed Goal work; preserve explicit Skill boundaries; and expose exactly which runtime can execute the next step.
From request to envelope and evidence
Continue the migration Goal across API, Web, and docs.
managed-goal
bundled-local-r0
plan_work, propose_support_consent, transition_support_consent, get_router_status
Scheduling requires verified Host capabilities.
Inspect every routing decision
MCP calls
Decision
Work graph
Audit trail
Evidence boundary
Discover what can actually run before routing
V2 separates installed metadata, Host exposure, authorization, policy eligibility, freshness, and risk. A local file or cached observation never becomes runtime authority by itself.
Read Runtime Capability DiscoveryChoose Single, Phased, or Managed Goal
The Router classifies dependencies, verification boundaries, resume needs, and Goal relation before selecting capabilities.
Apply Personal Routing Profiles
Workspace and personal preferences become deterministic intended routes without bypassing runtime discovery.
Preserve Explicit Skill Lock
Named SKILLs remain authoritative. Support outside the lock stays inactive until the user accepts it.
Activate only what current proof allows
Every selected capability keeps its exposure, authorization, policy, freshness, and risk boundary visible.
The result is inspectable before and after execution
The Router declares planned SKILL usage before work, records typed runtime results, and discloses actual usage at completion.
One Router, three work envelopes
Small work stays small, multi-stage work reroutes at phase gates, and durable dependency graphs bind to a Host-owned Goal without pretending local authority.
Compare routing envelopesWork shape decides the envelope
The Router does not force a Goal because a request is long. Dependencies, resume needs, risk, and verification boundaries drive the decision.
Request → Runtime Capability Discovery
Envelope → Single | Phased | Managed Goal
Result → plan, gate, evidence, disclosureDesigned for trustworthy outcomes
Own the route
Define repeatable personal or workspace Skill Trees without turning them into executable prompts.
Less context noise
Automatic routing stays minimal and does not add a consent ceremony when no SKILL was named.
Honest degradation
Unavailable Host or adapter capabilities return typed fallbacks instead of simulated success.
Auditable execution
Sanitized traces show what was planned, called, rejected, and actually used.
Evidence before claims
Contract fixtures prove compatibility. Runtime traces prove bundled execution. Real Model Evaluation requires fresh paired attempts, explicit quota authorization, review, and attestation.
Evidence boundary: 36 attempts and 42 model turns are historical beta.1 evidence. Personal Routing Profiles and beta.2 currently have deterministic contract, integration, security, and package evidence only.
# Inspect truthful runtime readiness
python plugins/workflow-skill-router/runtime/workflow_skill_router.pyz doctor
# Verify generated MCP contracts and V2 docs
node scripts/build-mcp-reference-data.mjs --check
python scripts/check-doc-parity.py
# Run repository tests
$env:PYTHONPATH = (Resolve-Path "packages/router-core/src").Path
python -m unittest discover -s packages/router-core/tests -v
python -m unittest discover -s tests -vView the evidence contractMake runtime-aware skill routing reusable
Install the Plugin + MCP runtime, or load the standalone SKILL when your Host supports instructions only.