Skip to content

Latest commit

 

History

History
38 lines (31 loc) · 2.27 KB

commit-namespace.md

File metadata and controls

38 lines (31 loc) · 2.27 KB

Commit-message issue namespaces (#483)

Commit messages in this repo cite bare #N references drawn from three unrelated numbering namespaces whose ranges overlap, so a bare #437 reads as whichever tracker the reader happens to have open. This has already manufactured false readings — see #483 for the 2026-07-24 near-miss (od:idle-cpu #437 grepping identically to GitHub issue #437) and the six work items that were reported to the operator against unrelated 2024-era PR numbers.

The convention

Namespace How to write it What it is
GitHub issue (this repo) bare #N (historical default) or gh#N The authoritative board (Project #28). A grep-checkable "did this land?"
Operator-debug series od:#N The od: scratch series (e.g. od:idle-cpu). MUST be qualified — a bare #N next to od: is the exact collision that bit us.
Session task-list card never cite to the operator An agent-side scratchpad the operator cannot open. A tracker nobody can open is not a tracker. Any operator-facing status MUST reference a GitHub issue, never a card number.

Rules (from #483, in force)

  1. A GitHub issue reference is a bare #N (or gh#N). Bare #N has always meant "GitHub issue in this repo" here; that stays true. Prefer gh#N when a message also touches another namespace, so a later grep cannot confuse them.
  2. An operator-debug reference MUST be od:#N — never od:label #N with a bare number, which is what grep misreads as a GitHub issue.
  3. Never cite a session task-card number to the operator. Reconcile work onto GitHub and report the GitHub number.
  4. Reconcile onto GitHub, never the reverse — Project #28 is authoritative (team-lead decision, 2026-07-24).

There is no lint for this, deliberately

This convention is carried by review, not by a gate. A hard gate rejecting a bare #N would wedge every commit and require repo-wide buy-in for a new syntax, and a lint that fires on the common, correct case trains everybody to ignore it — the "a guard that cries wolf gets switched off" trap this codebase warns about elsewhere. An earlier advisory script existed and was removed; nothing replaced it, and that is the settled position rather than a gap somebody should fill.