Guards

The harness cockpit, a viewer: the phase every agent is in (PHASES panel, and main and the running agents in the band above the prompt), the advisor calls, what the insights learned (per-phase evidence, session outcomes from /insights, goals, the calibration and its proposals, with buttons to run /insights, apply, or open a calibration goal), Cortex memory health and grooming, the context...

Two rules no settings hook enforces, refused at the tool call: wiki pages and their generated ADR mirrors are written only through the Cortex wiki tool, and every git worktree lives at /.claude/worktrees/ / (one created there is registered with disk_hygiene.py).

Automatic effort and model, and a quota that limits how many agents run, not how well one runs. Every request runs at the effort its grade calls for (zetetic-genius) and the main tool loop steps down the ladder. A subagent runs on the model its Agent call names, else the one its agent definition pins (read from the definition file, since agent.spawn carries only the call's), an inheriting type...

The learning loop: every main turn is kept in a ledger across sessions (phase, effort, steps, failed tool calls, agents, advisor calls), joined with the outcomes observed at tool.call (commits kept or reverted, test exits, PR verdicts and merges, no transcript) and the ones /insights writes, and the project's goals; probe sessions are counted and kept out; a phase whose evidence says so gets...