TaskCompleted Codex-review hook re-reviews the whole accumulated diff per task (repeats findings + diff-local false positives)

[LukeLIN-web]

TaskCompleted Codex-review hook re-reviews the whole accumulated diff per task, repeating findings and emitting diff-local false positives

What happens

When executing a /goal plan with the task list (TaskCreate / TaskUpdate), every time a task is marked completed a Codex review runs as a TaskCompleted hook and can block completion with CHANGES_REQUESTED. Observed across one plan execution (10 tasks, single shared working tree on main):

• task1 → 3 findings.
• task2 → 2 different findings — genuinely useful, caught a real design flaw (per-item backward vs. the intended allreduce-first DDP sync). This is the hook working well.
• task4 → 2 findings that were verbatim repeats of task1's findings about the same lines (already discussed/resolved on task1).

Root cause (apparent)

The hook reviews the entire accumulated working diff on each task completion, not the incremental delta introduced by that task. Consequences:

1. Repeated findings. Because tasks in a plan often touch the same files sequentially and share one working tree, a finding about code written in task1 is re-surfaced on task2, task4, … every subsequent completion. The reviewer does not appear to track which findings were already raised/addressed/refuted.
2. Diff-local false positives. The reviewer judges the hunk in isolation without surrounding file context. Example: it flagged open(os.path.join(args.out, "csv_probe_ids.json"), "w") as "writes before the output dir is created" — but os.makedirs(args.out, exist_ok=True) runs unconditionally ~20 lines above at the top of the same function. The reviewer can't see it, so the same false positive recurs.

Why it's a problem

• Each repeat costs a full Codex round-trip (latency + tokens) to re-litigate settled points.
• It pressures the agent to make redundant defensive edits (e.g. a second idempotent makedirs) just to silence a diff-local false positive, adding noise to the code.
• Genuinely new findings (like task2's) get diluted by the repeats, reducing signal.

Suggested fixes

• Scope the review to the task's incremental diff (e.g. diff vs. the tree state when the task was created/claimed), not the cumulative working tree.
• Persist a per-plan ledger of already-raised findings (by file+line+claim hash) and suppress / mark-as-prior on repeat, or pass prior findings + their resolutions into the reviewer's context.
• Give the reviewer enclosing-function (or full-file) context, not just the diff hunk, to cut diff-local false positives.
• Optionally let the agent record a refutation for a finding (with evidence) that carries forward, so a verified non-issue isn't re-raised.

Environment

• Workflow: /goal <plan.md> execution with TaskCreate/TaskUpdate + Codex TaskCompleted review hook.
• Single working tree, all tasks on main (no per-task worktree isolation).