feat(harness): thinking default-collapsed preference toggle via Leader+t (plan #742)

[btipling]

plan #742 — thinking default-collapsed toggle via Leader+T

Re-implementation of plan #742 against current main. The original implementation (#751) was stacked on plan/harness-keymap (parent plan #741) and was auto-closed without merging when #746's merge deleted that base branch. Those keymap seams are now on main, so this PR re-implements the plan directly onto main (Zig + docs only).

Fixes #742
Refs #741 (parent keymap)

Change (harness + docs only)

• thinking_collapse.zig (pure, host-testable) — shouldRenderFull now threads default_collapsed as an explicit boolean parameter:
  shouldRenderFull = operator_open OR (NOT default_collapsed AND isActiveTurnFull).
  Default ON relaxes the Busy pin; OFF restores today's behavior.
• ui/state.zig — thinking_default_collapsed = true, reset to ON in resetTranscriptScroll() (same lifetime/reset site as help_overlay_open / leader_armed). In-memory only; no session meta / protocol carrier.
• ui/thinking.zig — threads the live preference; the is_active "expanded = true" pin is gated (pinned_active = is_active and !default_collapsed), so when ON an active row's click routes through the normal set-toggle into thinking_open_l1 (expander works mid-Busy). No second if (is_active) pin left that ignores the bool.
• Keymap seams (all 4) — keymap.zig (Action.thinking_default_toggle + KEY_TABLE row thinking_default_toggle + a leader-pending t arm in match, mirroring the ? arm: handled, disarms, never inserts t), ui/keymap_dispatch.zig (Handlers.toggleThinkingDefault seam + a runAction arm that flips + disarms the leader), ui/help_overlay.zig rowChord case (exhaustive switch — compile-forced, Thinking default collapsed / Leader Space, then t), ui.zig seam + dispatch wiring.
• Docs (timeless, no issue references) — docs/harness-limits.md (keyboard row: Leader then t; Thinking collapse row: Busy pin now preference-gated), docs/agent-stream.md (qualified the absolute "active Busy turn keeps thinking fully expanded").

Verification (Zig gates, all green)

Gate	Result
zig fmt --check	clean
zig build test-rich	92/92 steps · 1299/1299 passed (keymap 39/39 incl. tests 5/5b/table-cap; thinking-collapse 13/13)
zig build test-rich-invariants	green
zig build harness -Doptimize=Debug	Wasm compile clean
build-harness CI	pending — will update on completion

No TS files touched → vitest / typecheck / DI-cost gates N/A (per the layer rule). No cap added/changed. No Production mutate; build-harness is the existing release gate.

Per the plan's testing lock (#1–6, 5b, 8; #7 compile-forced): tests 1–3 (both preference values + committed rows), 5/5b (leader+t → toggle, disarm, no prompt insert), table-within-cap.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 21, 2026 8:11am

[btipling]

Summary (re-implementation)

This is plan #742's fresh PR against main; the original stacked PR #751 was auto-closed without merging when parent #746's merge deleted its base branch. Reconciled here by replaying the single #742 delta onto the updated main (keymap seams from #746 are now on main). Zig + docs only.

Gates (all green): zig fmt --check clean · zig build test-rich 92/92 steps · 1299/1299 passed (keymap 39/39 incl. leader-t tests 5/5b + table-cap; thinking-collapse 13/13) · zig build test-rich-invariants green · zig build harness -Doptimize=Debug clean · build-harness CI green (run 32459134300, 2m41s, wasm32-freestanding release + artifact upload). PR MERGEABLE / CLEAN.

No TS changed → vitest/typecheck/DI-cost N/A. No cap added/changed; no Production mutate. Plan #742 marked IMPLEMENTED.

Next: adversarial review on this head 2c215e5.

[btipling]

Adversarial review — PR #754

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/thinking-default-collapsed · 9 files · Leader+t thinking default-collapsed (#742)
Lenses run: L1, L3, L5, L6, L8, L9 (skip: L2 no secrets/API; L4 no workflow; L7 no clone bind)
AGENTS.md read: yes · docs/feature-divide.md read (preference is Wasm in-memory; no meta / protocol / DOM)

The shape is right: shouldRenderFull = operator_open OR (NOT default_collapsed AND isActiveTurnFull) is the single policy, preference default ON, Busy pin gated in paint (pinned_active = is_active and !default_collapsed) so mid-Busy expand still writes thinking_open_l1. Leader+t is a short-circuit arm in match (same class as ?) — the KEY_TABLE row would never fire without it. fromDvui already has .t (Ctrl/Cmd+T stay reserved → .browser, disarm, no toggle). Dispatch flips then disarmLeader(). Reset on New/Clear/hydrate returns ON. Caps unchanged (KEYMAP_MAX 64, table still well under). Overlay exhaustive switch compiles. Docs qualify the old “Busy stays fully expanded” claim.

Tests lock both preference values + committed rows + leader+t / plain t / Ctrl+T reserved. No dual DOM cheatsheet.

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Nit	L8	thinking_collapse.zig file header still locks “Busy → every current-turn thinking row FULL” (#424). Product default is now the opposite.	Next collapse agent “restores” the Busy pin from the module comment and fights #742.	shouldRenderFull body + tests are the real API; header is stale prose.	high

Residual risk

• Test 5b names “disarms / never inserts t” but only asserts match outcome; disarm + handled-mark live in keymap_dispatch (not in test-rich).
• Overlay greys the new row unless leader_pending (same as leader+?) — technically correct, easy to misread as “unavailable”.
• Zig test-rich / wasm32 build-harness were not re-run in this review environment.

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; nits optional
• Do not persist the preference in session meta / a new inv_* in this PR
• Do not add a DOM window keydown for Leader+t
• Do not silently change KEYMAP_MAX / LEADER_WINDOW_MS

What was not attacked

Host zig build test-rich / wasm32 build-harness, live canvas (Leader+t mid-Busy expand/collapse, New session reset), Production harness artifact deploy.

[btipling]

Addressing adversarial-review Nit (L8) from the PASS WITH NOTES review on 2c215e5:

Change

native/harness/src/thinking_collapse.zig — reconciled the module header with the shipped policy (plan #742). The header locked the pre-#742 rule "Busy → every current-turn thinking row renders FULL" (#424). The product default is now the opposite (default_collapsed = true). Reworded the header to state the current rule:

• default collapsed: active-turn rows render compact unless the operator opened them;
• OFF preference restores the today Busy pin (active-turn rows FULL);
• on turn completion everything collapses regardless of preference.

The shouldRenderFull body + tests (the real API) were already correct — this was stale doc prose only.

New head: ec5f709 (author btipling)
Gates on the new head: zig fmt --check clean · zig build test-rich exit 0 · zig build test-rich-invariants exit 0 · zig build harness -Doptimize=Debug clean. Comment-only change (no code logic touched).

CI (zig → harness.wasm, Vercel) is re-running on the new head.

Per merge skill §0a, the head moved (2c215e5 → ec5f709) which invalidates the prior adversarial PASS — a fresh adversarial review on ec5f709 is required before /merge_pr #754.

[btipling]

Adversarial review — PR #754

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/thinking-default-collapsed · 9 files · Leader+t thinking default-collapsed (#742)
HEAD: ec5f709 (header-only follow-up after 2c215e5)
Lenses run: L1, L3, L5, L6, L8, L9 (skip: L2 no secrets/API; L4 no workflow; L7 no clone bind)
AGENTS.md read: yes · docs/feature-divide.md read (preference is Wasm in-memory; no meta / protocol / DOM)

Re-review. Prior Nit (L8: thinking_collapse.zig header still locked “Busy → FULL”) is fixed. Header now matches shouldRenderFull = operator_open OR (NOT default_collapsed AND isActiveTurnFull) and the product default ON.

Logic is unchanged from the previous attack and still holds: leader+t is a match short-circuit (same class as ?); Ctrl/Cmd+T stay reserved; dispatch flips then disarmLeader(); paint pinned_active is preference-gated so mid-Busy expand writes thinking_open_l1; reset on New/Clear/hydrate returns ON; overlay exhaustive switch compiles; caps unchanged; tests lock both preference values + leader+t / plain t / Ctrl+T reserved. No dual DOM cheatsheet.

Findings

None on ec5f709.

Residual risk

• Test 5b names “disarms / never inserts t” but only asserts match outcome; disarm + handled-mark live in keymap_dispatch (not in test-rich).
• Overlay greys the new row unless leader_pending (same as leader+?).
• Zig test-rich / wasm32 build-harness were not re-run here (comment-only delta).

Merge guidance

• PASS WITH NOTES: safe to merge from this attack
• Do not persist the preference in session meta / a new inv_* in this PR
• Do not add a DOM window keydown for Leader+t
• Do not silently change KEYMAP_MAX / LEADER_WINDOW_MS

What was not attacked

Host zig build test-rich / wasm32 build-harness, live canvas (Leader+t mid-Busy expand/collapse, New session reset), Production harness artifact deploy.