fix(ding): classify a used Claude pane instead of deferring forever

[schickling-assistant]

Fixes #26.

Delivery to a Claude pane never left Ambiguous, so every DING deferred indefinitely and messages only ever reached the inbox. Fixing that turned out to expose two further wrong-positives in the same routing path, which are fixed here rather than shipped.

Five commits: three behaviour, then a pure restructure and its whitespace cleanup.

3ec59fa	detect the composer, not the startup banner; parse an empty composer row
37c4467	classify the lowest composer on screen, not a preferred harness
04b600c	treat an in-flight Claude turn as blocked
d3f5fd7	split the composer into per-harness adapters (no behaviour change)
2cd0635	collapse blank-line runs the split left behind (whitespace only)

1. Harness detection depended on a banner that is often absent

classify_composer routed to the Claude branch on plain.contains("Claude Code v"). pty peek returns the viewport only — there is an explicit --full flag for scrollback — so the check is unreliable, and when it fails the function returns Ambiguous without ever looking at the composer.

It is worse than "the banner scrolls away". On a detached 80x24 pane the string Claude Code v does not appear at all: the boxed welcome panel renders the wordmark and a truncated model line instead. Measured on a live capture at boot, plain.contains("Claude Code v") was false. On a 170-column capture of a different pane the same string is present. Either way it cannot gate detection, so detection now uses the ruled ❯ composer, which is durable, and the banner check is gone entirely.

2. An empty composer row failed to parse

The composer row of an empty composer is ❯ + U+00A0 and nothing else. U+00A0 is Rust-whitespace, so rows.first()?.trim_end() erased the prompt and both strip_prefix calls failed, returning None. Strip the prompt, then trim.

placeholder also needed the empty case: Claude shows the rotating Try "…" placeholder only on an unused pane, and an empty composer is the stronger positive-empty proof, since a human draft cannot be empty.

3. Transcript text could outrank the live composer

Raised by automated review on this PR. Routing preferred any Claude-shaped text over the ANSI-detected Codex composer, so a Codex pane whose scrollback held a captured Claude screen was classified from the capture. Capturing and pasting pane text is routine.

This is a wrong positive, which makes it worse than the deferral bugs: the paste and the bare Return go to the pane's real bottom composer whichever text was read. A transcript row that looks empty lets the paste land in a human's live Codex draft; one holding the exact notice can satisfy both adjacent observations and submit that draft.

A pane has exactly one live composer and it is the lowest on screen, so compare positions instead of preferring a harness — scrollback is above the live composer by construction. The Codex locator reports a raw ANSI byte offset while the Claude locator works in stripped line indices; every Codex marker starts at an \x1b[ boundary, so stripping the prefix is faithful and its newline count is that composer's row.

4. An in-flight turn read as idle

Claude 2.1.220 ships no interrupt hint with its working status line. None of the nine strings interaction_blocked matched appear during a turn — across roughly seventy real captures spanning several live turns, every one scored zero. It is not width truncation: on a 170-column pane the line ✻ Frolicking… (3m 35s · ↓ 6.9k tokens) renders in 38 cells.

Until now that was harmless, because an active pane also has an empty composer and an empty composer failed to parse, so delivery deferred. Fixing defect 2 removes that accidental guard and leaves a working pane reading as positively idle — DING would paste and press Return in the middle of someone else's turn.

The predicate is derived from captures rather than assumed, and two observations shaped it:

• The leading glyph animates — ·, ✶, ✻, ✽ all observed — so it cannot be matched literally.
• The elapsed timer is absent from some active frames (✽ Schlepping…), so requiring it would fail open on exactly the frames that matter.

What separates the states is the ellipsis directly after a single gerund. Active is <glyph> <Gerund>…; finished is <glyph> <Past> for <n>s. That distinction is load-bearing in both directions: the finished line sits above every idle composer, so matching it too would have stopped delivery altogether rather than being merely conservative.

Restructure

d3f5fd7 is a pure move — no behaviour change, no test renamed, no assertion altered.

src/ding/
  mod.rs          transport, pending notices, DingConfig, run_ding, serve
  composer.rs     ComposerState, the router, strip_ansi, soft-wrap candidates
  harness/
    mod.rs        trait Harness + Located + the registry
    claude.rs     locate / classify / idle footer / blocked
    codex.rs      locate / classify / idle footer / blocked

The router is now max_by_key(row) over every registered harness, so the positional rule is one comparison rather than a special case per harness pair. Claude is registered last so an exact row tie still resolves to Claude, as the explicit comparison did.

Active-turn and modal detection moves to the harnesses, since those shapes are harness-specific chrome. The shared predicate was already Claude-flavoured — Create a plan?, Our systems are thinking a bit more, and Retry with a faster model were applied to Codex panes too — so it is duplicated verbatim into both adapters rather than narrowed, keeping the move behaviour-free. Narrowing the Codex side is separate work. Only the choice-menu heuristic, genuinely cross-harness, stays in composer.

INVARIANTS.md names tests by file path, so its seventeen src/ding.rs:: references are repointed to src/ding/mod.rs::. All tests stayed in that module: the screen fixtures are shared between the composer assertions and the transport ones.

Fail-closed

A human draft is still Changed, an active turn or modal still ExactBlocked via the harness's blocked predicate, an unrecognised screen still Ambiguous. Transcript text never outranks the live composer, and an in-flight turn is now genuinely blocked rather than accidentally unparseable.

One known limitation, pinned by a test rather than left implicit: permissions on matches only the bypass footer, so accept-edits and auto panes are never positively idle. With the empty-composer case added they return Changed where main returned Ambiguous; at the observed_poke entry point both are Deferred, so entry behaviour is unchanged.

Testing

• cargo test --lib --bins: 142 passed, 0 failed.
• nix flake check --print-build-logs: exit 0.
• Each new test was confirmed to fail without its fix, not merely to pass with it. Without the positional rule the empty transcript row yields EmptySafe; without the spinner predicate an active turn yields EmptySafe.
• Restructure gates: cargo test --lib -- --list is byte-identical before and after (142 tests); all 27 src/<file>.rs::<test> references in INVARIANTS.md resolve, checked both by grepping the named path for fn <name> and by running each with cargo test <full::path> -- --exact.
• End-to-end against a real Claude 2.1.220 pane in bypassPermissions, under a throwaway catalog and pty root: a [DING] new st2 message: … notice was typed into the composer and submitted, and the agent read, replied to, and archived the message. Re-run at 2cd0635 — the pushed head, after both the routing change and the module split — by polling the pane until the ❯ [DING] … line was caught in the composer, since the harness split rewrote the router that decides delivery.
• Negative control on the verbatim capture of that pane: at origin/main located_bottom_claude_composer returns None and classify_composer returns Ambiguous; on this branch, Some and EmptySafe. A mid-turn capture returns Changed here and Ambiguous on main — for the wrong reason, via the parse bug.
• The mode proven end-to-end is bypassPermissions only.

No captured pane text is committed; all fixtures are hand-written synthetic screens.

Posted on behalf of @schickling

field	value
agent_name	cl2-cedar
agent_session_id	e5217740-eef6-48eb-a794-3e5e11939e4d
agent_tool	Claude Code
agent_tool_version	2.1.220
agent_runtime	Claude Code 2.1.220
agent_model	claude-opus-5
runtime_profile	/nix/store/i8y8b542cyqi385ywcjw5fvsq24f75v4-coding-agent-runtime-profile/share/coding-agents/profile.json
skills_manifest	/nix/store/i81qxhzlrzcxrrdwpp6i8hagka2gby8y-agent-skills-corpus/share/agent-skills/manifest.json
worktree	dotfiles/schickling/2026-07-20-cos-misc
machine	dev3
tooling_profile	dotfiles@unknown-dirty

[schickling]

@codex

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e1e64b837a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".