plan-backlog

Backlog Breakdown

Decompose the approved plan into .codex-flow/TASKS.md. Each task is one Codex run and one review cycle.

Task format

## T1: <imperative title>
- Depends on: — | T<n>
- Files: <create/modify list>
- Requirements: <R-IDs covered>
- Steps: <concrete, file-level steps>
- Skills: <relevant skills from PLAN.md "Skills plan" (*Skills to use* + before-execution created skills), or — >
- Acceptance: <verifiable criteria for THIS task alone>
- Session: —
- Status: pending

Session and the transition log beneath Status are execution-time fields that the orchestrator fills in Phase 4. The backlog always writes - Session: — and - Status: pending; do not add transition lines while planning.

The Files: and Depends on: fields are also what scripts/task-waves.mjs reads to compute parallel execution waves — keep them accurate and file-level, not vague.

Sizing: size for one execution AND one review

A task is the unit Codex builds in one run and Claude verifies in one pass — size it for both:

• One reviewable concern per task: it should satisfy one clear acceptance check that a reviewer can confirm without holding several unrelated changes in their head. If review would have to juggle two unrelated concerns, that's two tasks.
• Bounded blast radius: prefer a diff a reviewer reads in a single sitting — roughly ≤ ~5 files / a few hundred changed lines. Bigger diffs review shallowly and risk the 64 KB diff cap truncating what the reviewer sees. Too big → split.
• ~5–30 min of Codex work as a secondary check: a task you can't describe in a few file-level steps is too big; a one-line tweak is usually too small to be its own review cycle — fold it in.
• Mandatory slice overflow: if context-slice.mjs reports mandatory slice content exceeds tokenBudget, split the oversized task in the backlog; never raise the slice budget.
• Self-sufficient: each task must be doable by a fresh Codex session from PLAN.md + the task text alone (a domain shift starts a new session with no prior memory). Put the context it needs in Steps and name the files to read — don't rely on "as we did in T2".