[Juliusolsson05]

Refs Juliusolsson05/agent-code#1327. App side: Juliusolsson05/agent-code#1343.

Plan: agent-code docs/plans/2026-09-26-codex-tall-draft.md.

Problem

classifyCodexComposerState searched for the › marker only 12 rows above the separator, so a draft of 13 or more composer rows read unknown. unknown is deliberately not occupancy. Agent Code therefore published provider-not-ready, and its own Enter appended to the human's draft and submitted both.

Evidence

New recording testing/fixtures/composer-0157/tall-draft-ctrlc.json: real codex-cli 0.157.1 at 100×30, idle, then a 20-line draft typed with Ctrl+J newlines (no Enter, no model request), then Ctrl+C twice.

• Codex does not scroll the marker away: the composer grows upward. At 20 lines, › is on row 7 and the draft fills rows 7–26.
• Replay: drafted while the draft is up to 12 rows, unknown from 13 composer rows (including the empty cursor line right after a Ctrl+J), empty after Ctrl+C.
• The whole stream was decoded before committing. The only identifying value, the session id in the exit text, is redacted at the same length. The fixture is pretty-printed like its sibling.

Change

The 12-row bounded search is unchanged; blank rows inside it are allowed, since a draft can hold blank lines. When it finds no marker, a second walk starts at the separator and follows only non-blank rows up to the top of the viewport. The one exception is a single trailing blank row, the empty cursor line a Ctrl+J leaves at the bottom of a draft (recorded). A blank row anywhere else, including inside the 12-row window, ends the walk, so a transcript › above a blank row is never reached.

Tests

On the recording:

• drafted at 11, 15 and 20 rows (red on main from 13 rows);
• empty after Ctrl+C.

Replays feed recorded chunks in batches of 50, with xterm drained between batches, and the every-height test advances one replay. No timeout is changed; ~0.8 s and ~1.2 s under load.

On constructed rows:

• negative first: a transcript › across a blank row, whether above the window or inside it at 1/2/3/4/12, stays unknown;
• two trailing blank rows stay unknown, while the single cursor line does not;
• the nearest marker wins;
• the exact 12-row edge: a marker 12 rows up behind a blank is drafted, 13 rows up is unknown;
• the corrected bound test, which used to expect unknown for 12 unbroken continuation rows. That expectation was the bug.

Mutants killed: bound 11 and 13, no blank stop, no cursor-line skip, a walk starting above the window, deleting the tall walk.

Review

A (Codex), B (Pi) and C (Codex) are all MERGE-READY after verification; C used its final pass. The dispositions are in the PR comment.

Verification

• src/parsers: 31 tests (23 in the recorded file). tsc --noEmit is clean.
• The package CI is green.

Residual

• A draft with a blank line anywhere between its marker and the separator, once it is taller than 12 rows, still reads unknown. So do a draft taller than the viewport (its marker scrolls off) and a draft ending in two or more blank lines. None is recorded. Agent Code's orchestration delivery refuses these, but the pane's own Enter does not (agent-code#1327 stays open).
• A dense markerless transcript directly above the footer would read drafted. Per Codex's source, the footer is drawn with the composer's marker, so this frame is not expected; it is unrecorded.

🤖 Generated with Claude Code

[Juliusolsson05]

Review disposition (final head 0e5ea5d)

Three reviewers: A Codex, B Pi (the GLM slot; opencode would not start on this machine), C Codex. Round 1 was FIX-BEFORE-MERGE from all three. Verification: A MERGE-READY (round 3), B MERGE-READY (round 2), C MERGE-READY (round 4, final pass).

Round 1

• A1 / C1, major (the walk past the 12-row bound started ABOVE the window, so a blank row inside it was never checked and a transcript › 14 rows up turned a markerless frame into drafted, i.e. occupancy): fixed in 26eccea.
  • The walk now starts at the separator, and every row up to the marker must be non-blank.
  • The one exception, from the recording, is the single empty cursor line a Ctrl+J leaves at the bottom of a draft.
  • Tests: blank rows inside the window at 1/2/3/4/12 stay unknown; two trailing blank rows stay unknown.
• A2 / B1 / (#1343 A, B) (replays read whatever frame existed after a 2 s wall-clock deadline, so under load a half-parsed frame read unknown): fixed. The replays feed recorded chunks in batches of 50 and wait for xterm to drain. The every-height test advances one replay label by label instead of replaying from scratch (a4842a2, 0e5ea5d). No timeout was widened: the tall tests take ~0.8 s and ~1.2 s at load 41–49. The first attempt (one chunk per drain, 0cd724a) ran past 5 s under load and was replaced.
• C, surviving mutants:
  • The exact 12-row edge (bound 11 / 13): fixed, with a marker exactly 12 and 13 rows up behind an interior blank.
  • The marker < 0 guard: equivalent, proven. Both searches walk upward from the separator and stop at the first marker, so the long walk can only find the nearest marker or stop and keep the bounded result.
• C, minor (the threshold is 13 composer rows, not 14): fixed in the comment and body.
• A, major (declined): a dense markerless transcript, i.e. 13+ unbroken non-blank rows topped by a transcript ›, directly above the status+hint footer, with NO composer marker on screen. Per Codex's source the footer is drawn by ChatComposer together with its › prompt. That shape occurs only when a draft taller than the viewport has scrolled its marker off, and then it is a real draft with no transcript visible. It is unrecorded and stated as a residual.
• A, suspicions (overlay options, Working rows, image rows above a remote marker): what the pure function accepts; no recording shows Codex painting those frames with this footer.
• B2, minor (the bound was pinned only to [5, 12]): fixed by the exact-edge tests.
• B3, minor (the fixture was minified): fixed. Pretty-printed like its sibling, same data.

Later passes

• Round 2 found only the replay timeout (above), fixed in a4842a2/0e5ea5d. C's round 3 still saw one 5.3 s run under added load; 0e5ea5d made the every-height test advance one replay (~620 chunks instead of ~1,450). C's final pass passed 23/23 on four loaded runs, with the tests at 984 ms and 1,382 ms.

🤖 Generated with Claude Code

[Juliusolsson05]

Manager correction: this merged while its branch was 6 commits behind main (#56 had landed), so its green CI did not cover the combination. There were no conflicts. CI on main's merge commit d42cc1da is being watched; any failure gets a fix-forward PR. Merges now go through a scripted gate that refuses a branch behind its base.