feat(harness): str_replace L2 visible EMBER/TEAL bands, pin-open during turn

[btipling]

Summary

Implements plan #715 (source #714).

#692 shipped str_replace L2 side boxes, but the operator still saw an ink-only old side, a collapsed expander, and a stalled canvas if they opened it. This PR:

• Paints old/new on ember_border / teal_border fills (visible against teal_bg) with muted rims
• Sets inner textLayout .background = false so the box fill shows through (house markdown/diff already did this; feat(harness): str_replace L2 TEAL/EMBER side bands #692 did not)
• Pins L0 open while Busy if the group has a str_replace (even while still running); pins L2 once detail exists; collapses when the turn ends (reuse thinking_collapse_state.isActiveTurnFull)
• Splits sides at decode (Item.str_replace_sides, slot-cached). Committed sides use cache_layout + mono addText — no per-frame addTextMixed on the bodies

No protocol bump. No TS emit/preview/cap change.

Files

• native/harness/src/rich/toolrun.zig — Item.str_replace_sides, decode assignment, toolRunL0PinnedOpen / strReplaceL2PinnedOpen + tests (decode encoded line + pin matrix)
• native/harness/src/ui/toolrun.zig — fill tokens, .background = false, cache_layout, pin (no HashMap write while pinned)
• docs/harness-limits.md — Tool-run L2 / Painter / Open state (timeless)

Verification

Gate	Result
tsc --noEmit	N/A — no TS changes (this workspace has no node_modules; no TS delta)
vitest run --changed	N/A — no TS test delta
zig build test-rich / test-rich-invariants / harness -Doptimize=Debug	CI only — no zig in this agent workspace
build-harness	required check after push
Caps	existing 4096 B/side + 20/6 unchanged
Cloud ops	N/A
Living docs	in this PR

New Zig tests (plan matrix): decode encoded str_replace stores sides / exec stays null; L0 pin on name-only; L2 pin requires active+name+detail; existing 8 splitter cases unchanged.

Test plan (operator, after Preview wasm)

• During an in-flight str_replace: L0 open while running; L2 open once detail exists; two visibly colored bands (not ink-only)
• Canvas stays interactive while L2 is open
• Turn complete → L2 closes; re-open still shows bands
• Click during Busy cannot keep it closed
• Delete → old band + visible empty TEAL strip
• exec / read_file L2 still one teal_text block; not auto-opened
• ~390 px readable, no horizontal gutter
• Refresh: restored str_replace collapsed; expanding still splits

Fixes #715
Fixes #714
Refs #691
Refs #687

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 20, 2026 3:51am

[btipling]

Plan #715 implemented on this branch. Zig gates are CI-only (build-harness pending). No merge.

Operator smoke after Preview wasm: in-flight str_replace should show visible EMBER/TEAL bands without a click; canvas should stay interactive; L2 should close when the turn ends.

[btipling]

Adversarial review — PR #716

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/str-replace-l2-fix @ 3e59a42 · 3 files · str_replace L2 visible bands + active-turn pin + decode-time sides
Lenses run: L1, L3, L5, L6, L8, L9 (skip L2: no trust-boundary / emit / redact / protocol change. skip L4: no workflow edit; zig → harness.wasm already ran test-rich + wasm. skip L7: no deploy bind)
AGENTS.md read: yes · docs/feature-divide.md yes (paint stays in-canvas; no DOM chat)

Findings

No surviving Blocker / Major after self-refutation. Host addToolStart still emits detail: '' while running, so L2 pin + cache_layout only engage on the complete (ok/fail) payload — sides_stable is true on first band paint. Pin helpers match thinking_collapse.State.isActiveTurnFull (cleared on Busy→ready/err). HashMap is not written while pinned, so turn-end collapses. Item.str_replace_sides is slot-cached at decode; slices die with Decoded. Palette stays on named tokens (ember_border / teal_border fills, muted rims, ember_text / teal_text ink). build-harness is green.

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L1	native/harness/src/ui/toolrun.zig · paintToolRun reuses (src, item_base+2) for the no-detail textLayout and the has-detail expander. This PR makes that type swap the default path: L0 is pinned open while str_replace is still running (static label), then the same id becomes an expander the frame detail lands.	In-flight str_replace: L0 open, L1 is textLayout at +2. Tool completes → expander mounts at the same (paintToolRun src, +2). First completed frame can steal persist/hit data from the textLayout slot (wrong caret, one-frame dead hit target on the L1 row). L2 body still paints (l2_expanded is computed before the widget).	Defender: mutually exclusive in one frame; dvui may reset on kind change; operator-opened L0 during a running exec already did this. Fails because both widgets take the caller src (not their own @src()) + the same extra, and this PR makes the swap the happy path for every in-flight str_replace, not a rare manual open. Distinct extras (keep +2 expander, static label on a free slot) close it.	medium
Minor	L8	docs/harness-limits.md Tool-run Two-level expand still ends “Clicking a row toggles; second click collapses; per-item isolation.” Header (default) still says default-collapsed with no exception. Open state (same table) says L0/L2 are pinned for in-turn str_replace and click cannot collapse.	Next harness edit follows Header / Two-level, treats the pin as a regression, and “fixes” it by writing the HashMap again — #714 reappears (collapsed / click-to-see).	Open state row is accurate. Fails as living-docs contradiction: two sentences in the same table describe opposite click policy. The toggle sentence needs the pin exception (or a pointer at Open state).	high
Nit	L6	native/harness/src/rich/toolrun.zig · ui id packing is unique… still uses slot_max = 4 while paint mounts +5..+8.	A later extra at +5 is already used; the test will not catch a new alias inside 5..8. 200·1024+8 < 1000003 still holds, so today’s extras do not collide.	#692 Nit fixed the paint comment (+0..+8); the packing test was not updated. Math is still safe. Documentation-of-the-lock only.	high

Residual risk

Attack did not breach a runtime path with current evidence.

• Band fills (ember_border #3a1e18 / teal_border #152528 on teal_bg #050a0c) are still a small lift. Visibility now hangs on fill-through (.background = false) plus ember_muted / teal_muted 1px rims. If Preview still reads ink-only, the plan’s escalation is ember_muted / teal_muted fill — still palette-only. Operator checklist in the PR body is unchecked.
• Side bodies are mono addText (plan-locked to kill the #692 stall). → / emoji / CJK in the replaced text will tofu; the else-branch still uses addTextMixed. Restoring mixed-face on sides reopens the stall.
• Empty-new TEAL strip still depends on dvui honoring min_size_content.h = fontMono().lineHeight() on a box whose child addTexts "". Splitter tests lock new.len==0; they do not paint.
• cache_layout is keyed on status != .running, not decode-cache hit. Host never streams str_replace detail while running, so the first band paint is already committed. A future host that grows a completed item’s detail in place would trip DVUI’s append-only-prefix assert.
• No 390 px wrap / no-h-gutter proof in CI. Preview wasm not exercised in this review.

Merge guidance

PASS WITH NOTES — safe to merge from this attack after the operator smoke (in-flight pin + visible bands, canvas stays interactive, turn-end collapse, click-during-Busy cannot keep it closed, delete strip, exec/read_file still one unpinned block). Minors are should-fix-soon (distinct +2 extras; docs toggle sentence). Nit optional.

zig → harness.wasm is green (test-rich includes src/rich/toolrun.zig). Do not treat this as --approve.

What was not attacked

Live Preview paint, restored-session L2 after refresh, Copy of the still-sentinel payload, dvui testing-backend layout of the empty new box / +2 type swap, full vitest (N/A — no TS), DO runner beyond the already-green build-harness job.

[btipling]

Landed the adversarial-review notes on f1d7e95:

• L1 — no-detail L1 label is extra +9; expander stays +2 so a running str_replace that later gains detail does not reuse persist
• L8 — Header / Two-level expand now point at the Busy-turn pin (Open state)
• L6 — packing test slot_max = 9 matches extras +0..+9

Zig gates are CI (build-harness). No merge.

[btipling]

Adversarial review — PR #716

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/str-replace-l2-fix @ f1d7e95 · 3 files · str_replace L2 visible bands + active-turn pin + decode-time sides (re-review after 3e59a42 notes)
Lenses run: L1, L3, L5, L6, L8, L9 (skip L2: no trust-boundary / emit / redact / protocol change. skip L4: no workflow edit; zig → harness.wasm green on this SHA — test-rich + wasm. skip L7: no deploy bind)
AGENTS.md read: yes · docs/feature-divide.md yes (paint stays in-canvas; no DOM chat)

Findings

No surviving Blocker / Major after self-refutation. Prior review on 3e59a42 (L1 +2 persist reuse, L8 Header/Two-level vs Open-state contradiction, L6 slot_max = 4) is closed on this SHA: no-detail L1 is extra +9, expander stays +2; Header / Two-level point at the Busy-turn pin; packing test slot_max = 9.

Host addToolStart still emits detail: '' while running, so L2 pin + cache_layout only engage on the complete (ok/fail) payload — sides_stable is true on first band paint (widgets +4/+6/+8 did not exist while running; DVUI byte_heights empty → no CacheLayoutError). Pin helpers match thinking_collapse.State.isActiveTurnFull (physical-slot range, cleared on Busy→ready/err, including a same-frame tryPromoteQueued). HashMap is not written while pinned, so turn-end collapses L2. Item.str_replace_sides is slot-cached at decode; slices die with Decoded. Palette stays on named tokens (ember_border / teal_border fills, muted rims, ember_text / teal_text ink). build-harness is green.

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Nit	L8	native/harness/src/ui/toolrun.zig · paintToolRun calls state.thinking_collapse_state.isActiveTurnFull with no call-site note that the module is turn-membership (physical slot range), not thinking-specific. Plan #715 locked “Comment at the call site. Do not rename.”	Next harness edit treats the thinking_* name as a layering bug and deletes the import / pin — #714 collapsed L0/L2 returns.	Defender: Open state in docs/harness-limits.md and toolRunL0PinnedOpen already state the policy. Fails only as a plan-locked comment, not a runtime hole.	high

Residual risk

Attack did not breach a runtime path with current evidence.

• Band fills (ember_border #3a1e18 / teal_border #152528 on teal_bg #050a0c) are still a small lift. Visibility now hangs on fill-through (.background = false) plus ember_muted / teal_muted 1px rims. If Preview still reads ink-only, the plan’s escalation is ember_muted / teal_muted fill — still palette-only. Operator checklist in the PR body is unchecked.
• Side bodies are mono addText (plan-locked to kill the #692 stall). → / emoji / CJK in the replaced text will tofu; the else-branch still uses addTextMixed. Restoring mixed-face on sides reopens the stall.
• Empty-new TEAL strip still depends on dvui honoring min_size_content.h = fontMono().lineHeight() on a box whose child addTexts "". Splitter tests lock new.len==0; they do not paint.
• cache_layout is keyed on status != .running, not decode-cache hit (markdown is stricter: hit only). Host never streams str_replace detail while running, and side widgets mount only once detail exists, so the first band paint has no prior byte_heights. A future host that grows a completed item’s detail in place, or a ring-index shift that reuses (src, extra) with different text, would logError(CacheLayoutError) and drop the height cache for a frame — not a panic.
• Click-during-pin: dvui expander toggles the passed pointer and paints the caret before the pin re-asserts expanded = true. List stays open the same frame; caret can flicker one frame. Same pattern as ui/thinking.zig. Click cannot keep it closed.
• A 200-item rolled group that contains one str_replace stays L0-open (all L1 rows) for the rest of the Busy turn. Plan-locked; L2 still only the str_replace item(s).
• No 390 px wrap / no-h-gutter proof in CI. Preview wasm not exercised in this review.

Merge guidance

PASS WITH NOTES — safe to merge from this attack after the operator smoke (in-flight pin + visible bands, canvas stays interactive, turn-end collapse, click-during-Busy cannot keep it closed, delete strip, exec/read_file still one unpinned block). Nit optional.

zig → harness.wasm is green on f1d7e95 (test-rich includes src/rich/toolrun.zig). Do not treat this as --approve.

What was not attacked

Live Preview paint, restored-session L2 after refresh, Copy of the still-sentinel payload, dvui testing-backend layout of the empty new box / click-during-pin caret, full vitest (N/A — no TS), DO runner beyond the already-green build-harness job.

[btipling]

Addressing the re-review (f1d7e95) Nit (L8): added a call-site comment at paintToolRun noting thinking_collapse_state.isActiveTurnFull is a turn-membership predicate (physical ring-slot range [active_turn_start_slot, ring_head)), reused here to pin the tool-run group open during its Busy turn — not thinking-specific. States explicitly NOT to rename/relocate it as a layering cleanup (Plan #715-locked). Commit da654c6.

No planning comment warranted this round — single Nit, obvious localized doc fix (skill §0). Zig gates local + build-harness CI green on da654c6. No merge.

[btipling]

Adversarial review — PR #716

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/str-replace-l2-fix @ da654c6 · 3 files · str_replace L2 visible bands + active-turn pin + decode-time sides (re-review after 3e59a42 / f1d7e95 notes)
Lenses run: L1, L3, L5, L6, L8, L9 (skip L2: no trust-boundary / emit / redact / protocol change. skip L4: no workflow edit; zig → harness.wasm green on this SHA. skip L7: no deploy bind)
AGENTS.md read: yes · docs/feature-divide.md yes (paint stays in-canvas; no DOM chat)

Findings

No surviving Blocker / Major / Minor / Nit after self-refutation.

Prior notes on this PR are closed on da654c6: L1 +2 persist reuse (static label is +9, expander stays +2); L8 Header / Two-level vs Open-state contradiction; L6 packing slot_max = 9; L8 call-site note that isActiveTurnFull is turn-membership, not thinking-only.

Traced on this SHA:

• Decode assigns Item.str_replace_sides from a real encoded line (unescape path); slices point into owned detail; exec stays null. Paint reads the field; it does not re-split.
• L0 pin is name-only (has_str_replace); L2 pin is active + str_replace + has_detail. HashMap is not written while pinned → turn-end collapse with empty maps.
• isActiveTurnFull(slot, messageHead(), RING_CAP) uses the physical slot from messageSlotAt, same as thinking.zig. Idle/slot == null → not pinned.
• Host still emits empty detail while running, so L2 widgets +4/+6/+8 first mount on the complete payload with cache_layout true and empty DVUI byte_heights.
• Fills are named tokens (ember_border / teal_border, muted rims, ember_text / teal_text ink). Inner textLayout .background = false. Side bodies are mono addText. EMBER is removed-line semantics (AGENTS exception), not error chrome.
• New test-rich rows cover decode sides, L0/L2 pin matrix, and extras +0..+9. build-harness is green.

Residual risk

Attack did not breach a runtime path with current evidence.

• Band fills (ember_border #3a1e18 / teal_border #152528 on teal_bg #050a0c) are still a small lift. Visibility hangs on fill-through plus 1px muted rims. If Preview still reads ink-only, the plan’s escalation is ember_muted / teal_muted fill — palette-only. Operator checklist in the PR body is unchecked.
• Side bodies are mono addText (plan-locked to kill the #692 stall). → / emoji / CJK in replaced text will tofu; the else-branch still uses addTextMixed. Restoring mixed-face on sides reopens the stall.
• Empty-new TEAL strip still depends on dvui honoring min_size_content.h = fontMono().lineHeight() on a box whose child addTexts "". Splitter tests lock new.len==0; they do not paint.
• Click-during-pin: expander toggles the pointer and paints the caret before the pin re-asserts expanded = true. List stays open the same frame; caret can flicker one frame. Same pattern as ui/thinking.zig. Click cannot keep it closed.
• A 200-item rolled group that contains one str_replace stays L0-open (all L1 rows) for the rest of the Busy turn. Plan-locked; L2 still only the str_replace item(s).
• cache_layout is keyed on status != .running, not decode-cache hit. A future host that grows a completed item’s detail in place would logError(CacheLayoutError) for a frame — not a panic.
• No 390 px wrap / no-h-gutter proof in CI. Preview wasm not exercised in this review.

Merge guidance

PASS WITH NOTES — safe to merge from this attack after the operator smoke (in-flight pin + visible bands, canvas stays interactive, turn-end collapse, click-during-Busy cannot keep it closed, delete strip, exec/read_file still one unpinned block). Nothing left in the table.

zig → harness.wasm is green on da654c6. Do not treat this as --approve.

What was not attacked

Live Preview paint, restored-session L2 after refresh, Copy of the still-sentinel payload, dvui testing-backend layout of the empty new box / click-during-pin caret, full vitest (N/A — no TS), DO runner beyond the already-green build-harness job.