frontier-orchestrator

Measure the exact working set in lines, estimate tokens at roughly 10 per line, and route to the cheapest model whose context window holds it under 75% utilization. Precise small edits go to spark or haiku, broad sweeps to gpt-5.4-mini, and harder slices to Sonnet 5 or gpt-5.4.

A cross-family oracle gate validates Claude-coordinated work using a GPT-family model with its own tools, so the reviewer's blind spots are uncorrelated with the workers'. Gates fire before final validation, after plan creation, on worker conflicts, and before irreversible changes.

The orchestrator's context window is reserved as headroom for coordination, so every file edit is delegated to a worker, even trivial one-liners. If a task involves code changes and zero workers were spawned, the orchestration has failed and work must be sliced and delegated.

Count lines of the exact files in the worker's brief with wc -l, multiply by about 10 tokens per line, add the brief and expected output, and keep the total under 75% of the model's window. Oversized tasks are split into smaller slices before escalating to a larger model.

Workers return a too_big, out_of_scope, or mismatch signal instead of compacting silently. The orchestrator then re-measures, splits the slice first, and only escalates to a larger model if the slice is genuinely not splittable.

Unplanned mode is allowed only for genuinely small tasks where the orchestrator maps the surface itself. If exploration passes about 50% of the orchestrator's context window before a plan exists, it must stop and demand a plan built with the user and oracle.