Definition of done

Status: Accepted

What “finished” means before a PR opens. The single checklist a change is held to, gathering the obligations scattered across this section into one place.

The principle

Production-ready always. Shipped code is finished — no deferral notes, no “come back to this”, no TODO (comment policy). If work remains, the work is not done, and the PR is not ready.

This is not a high bar arbitrarily set. A codebase where “done” means “mostly done” accumulates a second, invisible backlog inside the code itself, and that backlog is never paid down because it isn’t visible.

The checklist

A change is done when all of these hold:

Spec

Cites a spec identifier that exists on the spec’s main (GOV-R2, GOV-R3)
If it changed behaviour, the spec PR merged first (GOV-R4)
Citation is in a commit trailer and the PR body (GOV-R5)
No requirement ID appears in any code comment (GOV-R6)

Correctness

The cited requirements are actually satisfied — not approximately
New behaviour has tests; the must-cover paths are covered
Coverage on applicable code is 100% — the gate is green, and any exclusion is annotated in the source, not a silent gap (Q-R61, Q-R62)
Every new user-facing error carries a remedy (Q-R16 — it won’t compile otherwise)
No unwrap/expect/panic in non-test code (Q-R12)

Checks

rustfmt, strict clippy, arch tests, unit, golden, integration all pass (Q-R30)
No lint suppression added to src/ (Q-R13)
cargo-deny clean; no new advisory or disallowed licence (Q-R32)
No secret in any tracked file, tests included (Q-R33)

Documentation

Public items documented (Q-R20)
Comments obey the policy — why not what, 2–4 line blocks, no IDs
Repo-specific how is in .docs/, linked from code, not inline (Q-R10)

Cross-cutting

Interactive additions have a non-interactive equivalent (F1-R6)
User-facing output obeys the error model and accessibility
Platform-conditional code goes through the platform component, not scattered cfg! (ARCH-R35)

What “done” explicitly excludes

Stated so they can’t be smuggled in as done:

Not done	Why
“Works on my machine, untested elsewhere”	The platform matrix is a promise; a change that only works on one is unfinished
“Passing but with a suppressed lint”	The suppression is the unfinished part (`Q-R13`)
“Feature works, error path panics”	A panic is an unhandled error with no remedy (`Q-R12`, `G4`)
“Spec change to follow”	That’s drift with a promise attached — the exact thing governance prevents
“TODO for the edge case”	Shipped code is finished; handle it or scope it out explicitly

The reviewer’s job

The checklist is the author’s; the reviewer verifies it, and answers two questions CI cannot:

Does the code actually satisfy the cited requirement? CI confirms the citation resolves; only a human confirms the behaviour matches it.
Is anything here a judgment-rule violation? Redundant comments, premature abstraction, a runtime check where a type would do — none are machine-detectable, all are review-blocking.

A review that only re-runs what CI already ran adds nothing. The value is in the two questions above.