sirkirby’s routr helps a lead coding agent hand work to other coding tools using existing subscriptions. Jev reads the task brief and later checks the worker’s report. Local code ranks available subscription capacity; the lead chooses the worker’s model and verifies its work.

Ask about the brief before choosing a worker

  • The implementation sends fixed, versioned questions to TypeSafe’s System One endpoint, with jev-1.14.0 pinned in the provider registry.
  • Questions cover an unknown bug cause, coordinated changes, data-integrity hazards, missing context and whether the brief states a check.
  • A separate answer suggests basic, standard or strong work. Uncertain fact readings remain explicit for the lead to resolve.
  • Subscription ranking applies the user’s reserves and permitted work levels. Unknown remaining capacity stays unknown.
  • Cloudflare’s Clef and local Ollama are optional decision providers. Selected failures permit a backup; malformed answers and refused credentials stop that sequence.

A report check still needs an execution check

routr check reads the report against the brief. It flags missing verification, admitted gaps, skipped requirements, symptom patches and changes outside scope. A clean report still directs the lead to run its own check.

Typed answers constrain the response’s shape. They do not prove the task assessment is right or that a worker actually ran the tests it describes.

The evaluation changed the decision rule

The maintainer’s evidence log reports a revealing mismatch. On 36 real tasks, an earlier rule that chose the lower of two likely work levels matched labels 15 times. Choosing the most likely level matched 29 times, using the same saved answers. Shorter briefs had favored the lower-level rule, so routr changed its policy.

The same log reports that pooled Jev 1.14 runs matched 81.5 of 104 labeled levels, versus 82.7 for 1.13. Those are averages across repeated runs. The maintainer moved the pin after mixed gate results, citing real-outcome evidence rather than a clean evaluation win.

These are maintainer-reported results, not independently reproduced measurements. Labels and question tuning largely come from the builders; part of the measuring workbench is private. The project also reports that one delegated audit consumed more usage and took longer than its solo baseline.

Caveat: the README, evaluation notes and relevant implementation were inspected; neither the application nor live model calls were run. The evidence supports a testable orchestration pattern, not a general savings claim or security assurance.