Thanks for taking this on, @donle — the core idea is a real improvement and I want to land it. The loop-side FIFO is clean, the six exit-path queue-clears are right, and the repairedCalls.length === 0 && queue.length > 0 → continue patch is exactly the bug that would have orphaned queued steers otherwise. Test for "two steers consumed in order" is the one I would have asked for.

A few things to address before I merge:

1. Comments — code-as-documentation is the project rule. A handful of new comments restate what the code already says, which the tests/comment-policy.test.ts gate doesn't always catch but I do enforce manually. Specific spots to delete or compress to one why-line:

• src/cli/ui/App.tsx ~L2671-2683: the 5-line "Busy with an active turn: inject text as a mid-turn steer message…" block — the call site loop.steer(text) already documents itself. Delete the comment.
• src/cli/ui/App.tsx ~L3375: "The ghost preview from handleSubmit's steer branch already confirmed…" — delete; the empty branch is self-explanatory in context.
• src/loop.ts ~L702-704: "Consume one queued steer between model calls. This never mutates the currently running request…" — the loop is short enough to read; if anything stays, one line explaining why we consume between calls (not what).
• src/cli/ui/PromptInput.tsx L46-48 and src/cli/ui/ComposerArea.tsx L51: internal interface fields don't get JSDoc — the project rule is JSDoc only on src/index.ts re-exports. Names already say enough.
• src/cli/ui/PromptInput.tsx ~L133: the inputFrozen / inputActive derivation is self-evident from the names; the two-line block above is redundant.
• dashboard/src/panels/chat.ts ~L773-774: drop the "When busy, the server steers the text…" comment — the call site does the work.

2. The wrapper string is mine to author. This text:

[Mid-turn steer queued by the user. Do not treat this as a new task; use it only as additional guidance for the current task after completing the current step.]

is a prompt-engineering choice that affects model behavior across every steer, and I'd like to land that line myself in a follow-up after this PR merges. For this PR, please either:

• (a) Hoist the wrapper into a single exported constant (e.g. MID_TURN_STEER_WRAPPER) at the top of src/loop.ts so it's the one line I edit later, or
• (b) Drop the wrapper for now — inject the raw text as a user message and let me layer the wrapper in once we've decided whether it should be user-role text vs. a structured marker the model recognizes.

I lean (a). Either is fine.

3. pendingUser const-ification. Confirmed: on main, pendingUser is initialized null at step entry, never reassigned to non-null anywhere in the function, but still read at loop.ts:741 (allowEmpty: pendingUser !== null) and loop.ts:744. The let → const is behaviorally equivalent today, but the variable's existence implies there used to be a non-null path that got refactored out. Please either:

• Leave let and add a one-line // TODO(#nnnn): pendingUser is currently always null; either wire it back or remove after opening a tracking issue, or
• Remove pendingUser entirely from step() plus the call sites at L706/L716/L741/L744 (collapse to null literal) — clearer than leaving dead state.

4. Open a tracking issue for "mid-turn steer queue", and link it from the PR. The interaction primitive deserves a public design pointer for the next person who touches this surface — busy-input behavior is something I get asked about. One paragraph is enough; the PR description already has most of the material.

Once 1-4 are addressed I'll merge. CI side is already in great shape — just substance and policy.

Addressed the local review items:

• removed the redundant comments called out in the review
• hoisted the steer wrapper into MID_TURN_STEER_WRAPPER
• removed the always-null pendingUser path instead of leaving dead state

Opened follow-up tracking issue ^