Blog

4 min readfeature

The todo unification round: why you saw two to-do lists in plan mode

The plan-steps widget and the seeded Todos panel both mounted after approval. This round made them one list — verified live, with receipts.

  • #planning
  • #tui
  • #verification

The bug report was one sentence: "why am i seeing two to-do lists in the plan mode?" The honest answer took a root-cause pass to write — because both lists were real, both were ours, and each had been shipped by a different round without knowing the other existed.

What you were seeing

Plan Mode V2 (the planning round) ends in an approval dialog. On approval, the approved steps were rendered as a legacy plan-steps widget above the editor — [ ] step lines advanced by [DONE:n] tags. But gap P-06 had also shipped a plan→todo bridge: approval seeded the V2 todo overlay with the same steps. Two widgets, two keys, both aboveEditor, stacked. The seed hook shipped without retiring the old widget.

Three more defects rode along: the todo tool was still available while planning (nudging the model to build a second list mid-plan — Codex rejects update_plan in plan mode with an explicit error for exactly this reason); seeding appended instead of replacing, so approving plan B stacked it on plan A inside one panel; and the seeded panel only mounted after the model's first todo call — popping in late, next to the list it duplicated. [DONE:n] tags and todo flips never synced, so the footer chip could disagree with the visible list.

The rule the fix enforces

We checked what every serious agent does before writing a line: OpenAI Codex rejects the todo tool in plan mode at the tool layer; Gemini CLI's plan-mode allowlist excludes write_todos; Claude Code keeps the task list one live surface and never renders the plan as a second checklist. The consensus is one sentence: the plan is a user-approved proposal artifact, the todo list is live execution state — exactly one progress surface may be visible at a time, and the handoff happens once, atomically, at approval.

So now:

  • The seeded Todos panel is the ONE checklist above the editor. The legacy plan-steps widget is retired — it survives only as the fallback for a todo addon disabled via /addons, and the two can never be mounted in one process.
  • Approval replaces the foreground list: ids rewind, one replayable seed entry lands, and approving a second plan never stacks it on the first. The panel mounts at the approval moment, not after the model's first todo call.
  • The todo tool is off the plan-mode toolset while planning, and the overlay suspends while the PLAN chip is on — a leftover list never sits next to the plan's checklist. Your data is kept; approval or leaving plan mode brings it back.
  • Progress has two input paths that are the same list: a [DONE:n] tag flips the matching seeded todo; a hand-driven todo flip completes the matching plan step. The footer chip, the mission panel and the panel always agree — and /resume reconciles both directions before the first paint.

The approval dialog — the plan's checklist appears once, and nothing is seeded behind it

Execution: the seeded Todos panel is the one checklist above the editor, with the plan chip in the footer

Verified like a user

Eleven new tests pin the contract through widget-capture assertions (the plan-steps widget never mounts while the addon is live; approval mounts the seeded unown-todos; re-approval replaces; tags flip todos; todo flips complete steps; resume reconciles both directions). But the headline got more than that: a live PTY session drove the real TUI against the real model — /plan on, a real task, the approval dialog, execution, settle — and checked the literal pane text at every step. Then it checked the session file, because panes are timing-dependent and truth isn't: the seed entry count, the rewound ids, and the final agreement between the plan state and the todo state on every matching step. Fifteen checks, all green — including the one that matters most: no [ ] plan-steps row ever appears.

The lockstep evidence is the part we're proudest of. In the live run the model used both input paths — [DONE:1], then [DONE:2] [DONE:3], plus its own todo updates — and the plan state and the todo state advanced in perfect lockstep, entry for entry, ending in the completion gate firing with all steps done.

After [DONE:n]: the flipped rows in the overlay, one list, no second checklist

The settle: the completion toast over the one list

One honest note from the verification runs: the model is allowed to add tasks mid-execution (the todo list is live state, not a prison), so the overlay can legitimately show a task the footer chip doesn't count — the chip tracks the approved plan's steps. That's not a desync; that's the difference between an artifact and a workspace.

The receipts — findings F1–F6, milestones M1–M7, and the resumable session log — live in the session doc; the regression lane is scripts/tui-capture/verify_todo_unification.py.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts