plan: remove !busy guard from composer arrow-key history

[btipling]

Status: IMPLEMENTED

Implements: #701

Summary

PR #686 shipped composer ↑/↓ arrow-key history, but pressing ↑ or ↓ while the agent is busy (thinking/inference running) does nothing. The guard in ui.zig:647 reads:

if (!busy and state.queue_editing_index == null) {

The !busy guard blocks history navigation during inference. This was a reviewer recommendation (#686 R1) that the implementer accepted — the original intent per the plan (#667) was that Busy should NOT block history. This plan restores that original intent.

Root cause

native/harness/src/ui.zig:647 — the !busy guard. Two lines below it, historyApply(.older) already gates on in_hist or buf_empty — the only scenario where history is entered is when the composer is empty or already in history, both of which are safe during inference (history is a pure in-memory read, no bridge write, no alloc).

Fix

One-line removal: delete !busy and from the guard at ui.zig:647.

// Before:
if (!busy and state.queue_editing_index == null) {

// After:
if (state.queue_editing_index == null) {

The doc comment on lines 644–646 ("Skip history when busy…") also needs to be removed or updated.

Behavior after fix

Scenario	Before	After
Idle, empty composer, ↑	Loads newest user msg	Same (unchanged)
Idle, in history, ↑/↓	Walks history	Same (unchanged)
Idle, non-empty, not in history, ↑/↓	Passes through to textEntry	Same (unchanged)
Busy, empty composer, ↑	Blocked (↓ does nothing)	Loads newest user msg
Busy, in history, ↑/↓	Blocked	Walks history
Queue editing (any state), ↑/↓	Passes through to editor	Same (unchanged)

Why this is safe

1. No bridge write. historyApply is a pure read: it scans the in-memory message buffer, finds user messages by kind, and copies text into prompt_buf. No inv_* call, no host comms.
2. No alloc. The message buffer lives in bridge.RING_CAP stack space already held by ui.frame(). No heap allocation, no I/O.
3. No state mutation outside composer state. Only prompt_buf, history_index, history_draft_buf, history_draft_len, history_newest_fingerprint, and history_newest_fp_len are touched — all composer-local state fields.
4. The queue_editing_index guard stays. While editing a queued row, arrows still pass through to the text editor.
5. Hydrate drop still works during busy. The fingerprint-mismatch check (lines 539–563) runs unconditionally in frame(), outside the !busy guard. If a session hydrate replaces all rows while the operator is in history during a busy turn, prompt_buf is correctly restored to the saved draft and history_index cleared.

Layers

Layer	Change
Harness (Wasm)	ui.zig — remove !busy and from guard + update doc comment (~3 lines)
Vercel backend	No change
DOM	No change

No protocol bump, no bridge changes, no TS gates.

Ops & docs

• Cloud ops: N/A — no Production mutate
• Docs: docs/harness-limits.md keyboard table row updated: "Not active while Busy — arrows pass through to the text caret" → "Works while Busy (queue band visible, spinner active) — arrows navigate history during inference"
• Caps: No new caps. No existing cap changed.
• Zig gates only: zig fmt, zig build harness -Doptimize=Debug, zig build test-rich-invariants, zig build test-rich (existing composer_history suite), build-harness CI

Testing

Existing composer_history.test.zig (30 tests) covers all data paths in composer_history.zig and historyApply. The !busy guard is in ui.frame() which has no host-testable path (requires a running bridge + lifecycle). Manual smoke:

1. Start a long agent turn
2. While busy (spinner visible), press ↑ on empty composer → newest user message loads
3. Press ↑ again → older user message loads
4. Press ↓ → newer user message loads
5. Press ↓ past newest → saved draft restores
6. Press New/Clear while busy and in history → draft restores, history resets
7. Queue editing while busy → ↑/↓ still pass through to editor
8. Hydrate (load a session) while busy and in history → fingerprint mismatch drops history, draft restored

Review notes (2026-08-21)

Finding	Sev	Issue	Fix
1	Minor (L1)	Plan didn't note that original plan #667 intended Busy to NOT block history — this is a bug-fix, not a design change	Added to Summary and Design notes
2	Minor (L8)	Doc update text imprecise for keyboard table format	Replaced with full sentence matching existing row style

Reviewed: baseline verified against live main (4bc831b). Guard at L647 confirmed; historyApply confirmed pure read (zero inv_* calls); fingerprint hydrate drop at L539–563 confirmed runs unconditionally. No Blockers, no Majors.

Verdict: HANDOFF-READY — one-line guard removal, safe during inference.

Ship

Shipped in PR #707 (plan/composer-arrow-busy).