Torafirma Skill Router
A local-first, single-binary auto-router for discovering and loading large agent skill libraries without flooding model context.
Automatic routing — no user interaction required: once connected through MCP or the included interface skill, the agent searches a bounded local metadata catalogue, selects an explainable ranked candidate, exact-fetches the selected SHA-256 revision, and continues the user's original task.
Skill Router keeps one small interface skill in context and leaves the full library on disk. Version 1.1.0 formalizes the ranking function, emits replayable decision records, pins skill bodies by revision and catalogue generation, and separates read-only consumers from privileged publishers.
What changed in 1.1.0 source
- Public ranking policy: tf.skillrouter.hybrid-lexical.v1.
- Full score decomposition in search output and append-only decision receipts.
- SHA-256 revisionid for exact skill-body identity.
- Deterministic cataloggeneration for the ranked catalogue snapshot.
- Fail-closed search-to-fetch verification.
- Pin-on-fetch semantics; no implicit hot reload during an active task.
- Read-only consumer catalogue connections.
- Separate writable telemetry database.
- Operator-only registration, indexing, deprecation, and archival.
- Revision-pinned MCP tools and resources.
- 61 deterministic regression and contract tests.
- End-to-end exact-fetch and eight-worker telemetry concurrency smoke test.
- Linux and Windows CI workflow.
The packaged 1.0.0 Windows artifact remains available while the 1.1.0 source change is reviewed and packaged.
One executable, four interfaces
| Interface | Intended use | Catalogue access | |---|---|---| | MCP stdio | Automatic agent routing | Consumer/read-only by default | | CLI | Automation, diagnostics, search and exact fetch | Consumer by default; explicit operator commands | | Loopback HTTP | Trusted local integrations | Consumer/read-only by default | | Interactive shell | Human administration | Operator/read-write |
Running skillrouter with no arguments opens the operator shell. Running skillrouter --help prints scripting usage.
Why it exists
Large agent installations can contain hundreds or thousands of specialized skills. Injecting every description and body into every prompt is expensive and noisy. Skill Router reverses that model:
- The user gives the agent an ordinary task.
- The agent invokes the small skill-router interface internally.
- The router searches compact indexed metadata.
- The agent receives bounded ranked candidates with score explanations.
- The agent selects one candidate and exact-fetches its returned revision and generation.
- Only that verified SKILL.md enters context.
- Routing decisions and fetch receipts are recorded in separate telemetry storage.
Ranking contract
The current policy is deterministic and lexical: