Tool-call output collapse: verify landed defaults, fix details-hint chord, decide remaining gaps

[Hmbown]

Problem

The original complaint: tool-call output showed full command + all stdout regardless of outcome, so successful runs were ~80% scaffolding noise; the ask was collapsed intent-first success rendering, auto-expanded failures, configurable verbosity, rollups for batched calls, and removal of the then-broken Alt+V affordance. Since filing, most of that has landed (see evidence). What remains is one verified inconsistency — the in-transcript details hints advertise a chord (v) that does not open details, the inverse of the original dead-Alt+V complaint — plus two explicit product decisions and a close-out audit of the original checkboxes.

Current evidence

Verified at v0.9.1 (2026-07-18) — landed behavior:

• Collapse mode + config: ToolCollapseMode (Compact default / Expanded / Calm) at crates/tui/src/tui/app.rs:510; from_setting accepts collapsed/collapse as aliases for canonical compact (comment cites this issue; landed via feat(settings): accept collapsed as alias for default tool-collapse mode (#3256) #3695, commit 7a70455 per maintainer comment). Setting tool_collapse_mode defaults to "compact" (crates/tui/src/settings.rs:221,393), settable via tool_collapse / tool_collapse_mode / collapse (settings.rs:759) with round-trip tests (settings.rs:2102).
• Success collapse: a successful shell call renders as a single header line in Live mode (+time only if ≥1s) — ExecCell::render, crates/tui/src/tui/history.rs:722–739. Non-failed tool cells truncate to 2 lines + a details affordance when show_tool_details is off (history.rs:290–299; show_tool_details defaults to false, settings.rs:407, keys show_tool_details/tool_details settings.rs:810).
• Failure verbosity: the 2-line and calm-mode truncations skip failed cells (!cell.is_failed() guards, history.rs:290/301), and failed exec cells fall through to command + output rendering (history.rs:741–812). Live output is capped at TOOL_OUTPUT_LINE_LIMIT = 6 with an explicit "N lines omitted; … full output" marker (crates/tui/src/tui/history/tool_output.rs:385–397); the pager holds the full text.
• Batched rollups: crates/tui/src/tui/history/tool_run.rs groups contiguous successful low-risk tool cells into ToolRuns with summaries like "Explored 3 files, 2 searches: read_file, grep_files, ran 1 command" (tool_run_summary, tool_run.rs:258). Failed/running/patch/review/diff/plan cells split runs (is_collapsible_tool_cell, tool_run.rs:132). Consumed by crates/tui/src/tui/widgets/mod.rs:209,585, gated by App::tool_collapse_active (app.rs:4290) with run threshold hardcoded to 3 (app.rs:3145).
• Alt+V is no longer dead: is_tool_details_shortcut (crates/tui/src/tui/shell_key_routing.rs:115 — Alt+V, or macOS Option+V legacy √; plain v/V never match, tested at shell_key_routing.rs:158–169) routes to open_tool_details_pager (crates/tui/src/tui/ui/activity_detail.rs:530) from the main input loop (crates/tui/src/tui/ui.rs:5052). The keybinding catalog documents Alt+V with a guard test tool_details_documents_alt_v_only_never_bare_v (crates/tui/src/tui/keybindings.rs:214,499).

Verified remaining gap:

• The transcript affordance hints are built from tool_details_shortcut_label() which returns "v" (crates/tui/src/tui/key_shortcuts.rs:61–63), so cells show "v opens details" / "v opens full command" / "v opens full output" (history.rs:294,305,1710; tool_output.rs:392) — but in the transcript/composer context bare v only types v (by design, TUI-DOG-002). The advertised chord does not work; the working chord (Alt+V, ⌥V on macOS) is not advertised.

Not implemented from the original ask:

• An "intent" line (one line saying why the call was made). The first line today is tool name + arg/command summary (summarize_tool_args, tool_output.rs:93; command_header_summary, history.rs:1726), not model-stated intent.
• The rollup threshold (min run size 3) is not exposed as a setting.

Scope

1. Fix the details-hint chord: make tool_details_shortcut_label() (key_shortcuts.rs:61) return the chord that actually opens details in the transcript (Alt+V, platform-aware for ⌥V), or source it from the keybinding catalog entry (keybindings.rs:214). Update the existing hint assertion at crates/tui/src/tui/ui/tests.rs:3905 and any history/tests.rs lines that snapshot affordance text.
2. Add a regression test tying the hint label to the working binding (alongside tool_details_documents_alt_v_only_never_bare_v, keybindings.rs:499) so hint and handler cannot drift again.
3. Decision (maintainer): is tool-name + arg/command summary sufficient for the "intent" line, or should an explicit intent source be added? Record the decision on this issue; implement only if accepted.
4. Decision (maintainer): expose tool_collapse_threshold (app.rs:3145) as a setting, or document it as fixed. Record and implement/close accordingly.
5. Close-out audit: re-verify each original checkbox against the anchors above and close this issue once 1–2 land and 3–4 are recorded.

Key files

• crates/tui/src/tui/key_shortcuts.rs
• crates/tui/src/tui/shell_key_routing.rs
• crates/tui/src/tui/keybindings.rs
• crates/tui/src/tui/history.rs
• crates/tui/src/tui/history/tool_output.rs
• crates/tui/src/tui/history/tool_run.rs
• crates/tui/src/settings.rs
• crates/tui/src/tui/app.rs

Acceptance criteria

• Every visible details affordance names a chord that actually opens tool details in the focus state where the hint is rendered (no "v opens …" text where bare v types v).
• A test fails if the hint label and the details keybinding diverge.
• Collapsed-by-default success rendering, failure verbosity, tool_collapse config (including collapsed/collapse aliases), and run rollups are confirmed against the anchors above — re-verified, not reimplemented.
• Decisions on the intent line and threshold setting are recorded on this issue (implemented or explicitly descoped).
• Issue closed with a comment mapping each original checkbox to its landing commit/anchor.

Verification

cargo test -p codewhale-tui --bin codewhale-tui --locked history
cargo test -p codewhale-tui --bin codewhale-tui --locked keybindings
cargo test -p codewhale-tui --bin codewhale-tui --locked shell_key_routing
cargo test -p codewhale-tui --bin codewhale-tui --locked settings
cargo fmt

Manual: run the TUI, execute a failing and a succeeding shell call, confirm the truncated cell's hint chord opens the details pager as advertised.

Out of scope

• Restructuring the history renderer modules (v0.9.2: Split TUI history renderers into focused modules #3308 owns that; this issue only changes hint text/labels and records decisions).
• Changing collapse/rollup grouping rules or the failure-split behavior.
• New keybindings beyond making the advertised chord truthful.

Related

• feat(settings): accept collapsed as alias for default tool-collapse mode (#3256) #3695 (landed the collapsed/collapse alias + settings coverage, credited to @findshan).
• v0.9.2: Split TUI history renderers into focused modules #3308 (history renderer split; sequence to avoid conflicts in history.rs).

Triage note: body restructured for agent execution on 2026-07-18; prior comment refinements folded in. Original wording preserved in edit history.

[Hmbown]

Landed #3695 (7a70455) to accept collapsed/collapse as aliases for the canonical compact tool-collapse mode, with help text and round-trip settings coverage.

I am leaving this issue open because the original request also discusses batched-call rollups / broader display behavior, but the naming/config gap is now covered. Thanks @findshan for the precise follow-through here.