agent-discipline

A methodology that lives only in a document is one the model is free to ignore. agent-discipline closes that gap by translating the Harness Engineering Methodology spec into three runnable capabilities: an auto-injected prompt section (order -90) that constrains the session lifecycle and core invariants; a scaffolding tool, agent_discipline_init, that seeds twelve artifact templates — entry instructions, feature list, progress log, verification scripts, naming registry, cleanliness checklist, and more — into a target repo while skipping files that already exist; and a read-only auditor, agent_discipline_audit, that runs six CRITICAL checks, scores five subsystems on a 1-5 scale, flags the weakest link, and returns a pass/fail verdict. Together they turn the spec from wall-decal advice into concrete toolchain checkpoints.
The deeper shift is the so-called Form C: the feature state machine is promoted from hand-edited JSON to a ctx.features event-sourced service. WIP=1 violations, marking a feature passing without verification evidence, and illegal state transitions — all of which in Form B could only be caught by audit spot-checks — are now structurally rejected as corrupt by the runtime. The accompanying get_feature / update_feature / verify_feature tools replace manual JSON edits, and verify runs the feature's verification steps first, committing complete only on success, making the verify-then-record order a non-optional pipeline.
The plugin is aimed at independent developers or small teams already working with DeepSeek Harness for engineering-grade development who have a methodology spec and want it to actually hold. If you are tired of a beautifully written document that the model still treats as optional, agent-discipline fills the execution gap between the spec and the toolchain.