[Juliusolsson05]

Fixes #1291.

Found by the Stage 3 C1 hunt. Plan: docs/plans/2026-09-25-claude-rollback-attribute-composer.md (first commit). This PR needs claude-code-headless#63 (merged, 64fd0eaf), and the submodule is bumped to it.

Evidence

The paste-debug corpus (1,082 journals) holds 4 rollback-exhausted {presses: 64}, 1 rollback-cleared and 7 rollback-unobserved. After a kill empties Claude's composer, Claude repaints placeholder text (a prompt suggestion, or a hint). The text-only read calls it a draft, so every kill looks like a failure and the prompt is yanked back.

The exhausted journals carry no screen, so the repainted placeholder is the explanation that fits them (one prompt was a typical suggestion, "yes fix all 9"), not a recorded observation.

What merges

• A live read. ClaudeSession.readComposer() returns the screen and the composer cell attributes (getComposerAttributes(), from style(dispatch): pad content from status index #63) from the LIVE terminal buffer in one call. It never uses the throttled per-frame state, which lags a 25 ms kill loop (review round 1).
• The classifier (classifyRollbackComposer):
  • Text-only base, fail closed. Its empty is trusted for a bare marker.
  • Overruled to empty only when the live attributes show a painted placeholder: dim cells, no plain cells, and an attribute-aware read of empty (which also checks continuation rows).
  • Overruled to drafted when text-only says empty for an allowlisted hint ("Press up to edit") but the live cells are plain (review round 2). A kill clears only the current visual row, so typed text reading like a hint can remain, and a retry would append to it.
  • A leftover character under the inverse cursor stays drafted.
• The observe → kill → yank structure is unchanged.

Tests

• Rollback cases: a dim suggestion ends rollback-cleared after one kill (red on main), an inverse leftover and a typed known-hint row are not cleared, and a bare marker with no attributes clears.
• A real ClaudeCodeHeadless-painted frame test: dim suggestion, typed text, a bare marker, a dim hint, and a typed hint.
• A dim marker row with a typed continuation row stays drafted.
• ClaudeSession.readComposer: null without a headless, and screen and attributes read together.

The Claude runtime suite passes, and npx tsc -b is clean.

Residual

An all-dim draft (for example native voice interim text) could look like a placeholder. No recorded rollback frame shows it; tracked with #1294.

🤖 Generated with Claude Code

[Juliusolsson05]

Round 2 disposition (final round), head ab9ab02

• A + B (major): typed text matching a known hint (❯ Press up to edit, plain cells) was classified as cleared. Fixed. When live attributes show plain (typed) cells on the marker row, a text-only empty becomes drafted. The bare marker and the painted dim hint stay empty. Fail-first: the real-ClaudeCodeHeadless painted-frame assertion and a rollback test (rollback-cleared / retrySafe) were red, now green.
• B (minor): the real adapter was untested. ClaudeSession.readComposer() now has a test: null without a headless, and screen and attributes from the live headless in one call.
• A/B surviving mutation (continuation guard): pinned. Attributes sample only the marker row, so a dim marker row with text on a continuation row must stay drafted.
• A suspicion (an all-dim draft, e.g. native voice interim text): not elevated. No recorded rollback frame shows it; it stays with Claude composer gate reads cached/attribute state that can be stale or misclassify a draft #1294.

205 Claude runtime tests pass; npx tsc -b is clean. No round 3 per §7.3. Merge order is unchanged: claude-code-headless #63 first, then the bump here.

🤖 Generated with Claude Code

[Juliusolsson05]

Gate check before merge: head 4e453b59 contains main 7f273e06, and quality-gate and minimum-node-fixture-gate passed (the earlier failure was the #1261 palette flake, fixed by #1300). Three reviewers; disposition above. The body is updated to the merged behaviour. Merging.

🤖 Generated with Claude Code