Camus Loop Studio

Run settings: decisions, recorded

maker (words + Flexible Build; proof gate uses Claude) reviewer (words + Flexible Build; proof gate uses Codex) reviewer effort (supported seats; proof gate pins this every round) review rounds (1–6) depth (words lanes)

Saved locally under ~/.camus; tracked public defaults stay unchanged.

MODEL QUALIFICATION

Qualification proves that this machine can reach the exact maker or reviewer tuple and parse its output. It is not a quality endorsement or production-review admission. Launch checks it again.

Connect a model server — declarations only

Choose a starter, inspect the exact JSON, replace every placeholder, and save it to local operator state. Saving makes a tuple visible, not trusted; qualification is a separate explicit provider call.

1 Describe the job — brief it like you'd brief a strategist 2 Define what must be true for you to trust the result

This is what the reviewer holds the work to. Change it and any earlier review of this work no longer counts.

add a clause:

3 Choose the kind of work

4 Choose maker and reviewer, then run

maker reviewer

One model makes the work. A separate reviewer tries to break it; independence is earned only when the recorded identities support it.

Auto-route from calibrated A/B evidence Task class

Off. The pairing shown above stays in control.

Build execution Target repository

Flexible Build uses both selected seats. Start from a clean repository. Changes stay in a separate worktree; review is advisory and nothing is automatically committed, merged, or published. Stop and resume preserve the same candidate, contract, and budget usage.

Verify command (optional)

Flexible mode: empty means untested. Supply your test command to verify the candidate. It executes locally with credential environment variables removed, not in an OS sandbox; use trusted projects. The proof gate can detect the stack. No $, backticks, quotes, backslashes, or newlines: put complex commands in a script.

Maker harness

Camus offers a native harness only after a spend-free local version and artifact check. Provider/model qualification is separate; credentials remain in the host-owned one-model gateway.

Recovery and budget limits

This check is safe to repeat after fixes or interruption

Without this authorization, an additional verification pauses for permission. No changed contract or model is allowed on resume. These are global run limits, not per-round allowances.

Model calls Maker steps File/tool actions

Repairs Retries Native recoveries Token budget (0 = raw only)

Active time (ms) Per-call timeout (ms) Inactivity (ms, 0 = off)

Token limits reserve capacity before a call; they are not provider-enforced billing caps. Unreported usage remains unknown. Inactivity detection is off by default: a quiet model may still be working.

After a run, inspect its receipt without a provider call: camus build --inspect RUN_ID.

Grounding Ground in Hivemind

Drafts can lean on Hivemind, Myosin's private research knowledge (on staging today).

External publication Publish the completed artifact to Hivemind

Off by default. Turning this on is explicit consent for the finished words artifact to leave this machine.

Depth

Quick, from Settings. Change it there; it applies to the next run.

Judge the artifact, not its maker.

Deliverable