send-message.sh: locate_input_box() can false-positive 'delivered' on a pane with no input box and two decoy rules #1364

Open
opened 2026-08-21 22:14:56 +00:00 by veronica · 0 comments
Member

Follow-up to #1362 / PR #1363, raised by rev-code-01 in review 255 as an adversarial addition (Should Fix, not a blocker).

What

PR #1363 replaces send-message.sh's Claude-only prompt-glyph probe with locate_input_box(), which falls back to locating the pi input box by its two horizontal rules. That kills the systemic pi false-negative (#1362).

The fallback anchors on the last two rule lines in the capture. A pane whose recent output happens to contain two rule-shaped lines, and which has no input box at all, therefore resolves to a range between two decoy rules.

Measured

rev-code-01, 2026-08-21, against dc6db5f1: a decoy-only pane (no input box, two output rules) returned rc=0 "delivered" where pristine origin/next returned unconfirmed.

The anchor was separately verified correct against decoy rules sitting above a real box — that case picks the box. The window is specifically the no-box case.

Why it is a follow-up and not a blocker on #1363

The trade is right. Pre-fix, every send to every pi seat reported may-be-UNDELIVERED (4/4 in one day, measured by peggy). Post-fix, a narrow shape can report delivered when the pane has no input box. A false negative on every pi send is a larger and more corrosive failure than a false positive on a pane shape nobody has observed in the fleet: D19 records what institutionalized distrust of the transport already cost.

Fix direction

Bottom-anchor hardening: require the bottom rule to fall within the last few lines of the capture. An input box is at the bottom of a pane by construction; scrollback decoys are not. That closes most of the window without reintroducing the harness-specific probe.

Register

docs/FRAMEWORK-DEFECTS.md D49.

Follow-up to #1362 / PR #1363, raised by `rev-code-01` in review 255 as an adversarial addition (Should Fix, not a blocker). ## What PR #1363 replaces `send-message.sh`'s Claude-only prompt-glyph probe with `locate_input_box()`, which falls back to locating the pi input box by its two horizontal `─` rules. That kills the systemic pi false-negative (#1362). The fallback anchors on the **last two** rule lines in the capture. A pane whose recent output happens to contain two rule-shaped lines, and which has **no** input box at all, therefore resolves to a range between two decoy rules. ## Measured `rev-code-01`, 2026-08-21, against `dc6db5f1`: a decoy-only pane (no input box, two output rules) returned **rc=0 "delivered"** where pristine `origin/next` returned unconfirmed. The anchor was separately verified correct against decoy rules sitting **above a real box** — that case picks the box. The window is specifically the no-box case. ## Why it is a follow-up and not a blocker on #1363 The trade is right. Pre-fix, every send to every pi seat reported may-be-UNDELIVERED (4/4 in one day, measured by peggy). Post-fix, a narrow shape can report delivered when the pane has no input box. A false negative on every pi send is a larger and more corrosive failure than a false positive on a pane shape nobody has observed in the fleet: D19 records what institutionalized distrust of the transport already cost. ## Fix direction Bottom-anchor hardening: require the bottom rule to fall within the last few lines of the capture. An input box is at the bottom of a pane by construction; scrollback decoys are not. That closes most of the window without reintroducing the harness-specific probe. ## Register `docs/FRAMEWORK-DEFECTS.md` D49.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: mosaicstack/stack#1364