Implement code test-first via the red→green→refactor loop. Use in the implement stage for every issue. Never write implementation before a failing test; never refactor on red. The generator's core working rhythm.
Entropy garbage-collection. Periodically scan for drift — stale docs, dead "always-loaded" context, smells, mismatched conventions — and open small targeted fixes. Use to keep the smart zone smart and the repo coherent for future agent runs. Runs orthogonally to the main lifecycle.
Drive a long-running "work → check → repeat" loop until an exit condition is met, max iterations hit, or a human-judgment flag is raised. Use for autonomous long tasks where an agent does work and an evaluator decides whether it passes (e.g. "keep fixing until all reviewers pass"). Backed by bin/ralph.sh.
Grade work against the plan's acceptance criteria from a SEPARATE context — tiered by risk (L0 self-check, L1 lightweight reviewer [default], L2 full evaluator). Use in the evaluate stage. The anti-self-praise firewall — the agent that built the work must never be the one that grades it. Returns pass/fail per criterion with evidence.
View and advance the lifecycle stage of every requirement. Main-agent only — the single authority that moves a plan through intake→prd→issues→implement⇄evaluate→checkpoint→done. Use to see what stage work is in, route simple vs complex requests, and promote a plan (no sub-agent may self-promote).
Write a committed handoff checkpoint so a fresh session can resume with zero loss. Use when crossing the 40% context line, on every stage transition, and at session end. Writes .trace/checkpoints/<plan>-<seq>.md while the agent is still sharp.
Cut a versioned release of harness-mini — bump VERSION, roll the CHANGELOG, tag, and publish a GitHub release. Use when shipping a new version. Wraps `bin/harness.sh release`; covers the human-judgment steps (semver choice, changelog curation) the script can't make.
The horizontal+vertical coding constraint. Use in the implement stage to decide build order and enforce layering. Vertical = one feature end-to-end as a thin walking skeleton first; Horizontal = forward-only dependencies through a fixed layer stack. Read before writing code.