Phase 3: pausable command lifecycle (pause / resume / cancel)

[aboimpinto]

Phase 3 — Pausable Command Lifecycle (Pause / Resume / Cancel)

This PR builds on the custom slash command frontmatter from Phase 1 (PR #2326) and the hook gate from Phase 2. It adds pausable: true lifecycle support for custom slash commands: pause with ESC, type other messages while paused, resume with "continue"/"resume", and cancel with ESC twice.

Architecture

Model behavior is handled by centralized LLM evaluation. A single add_paused_evaluation_note() helper appends an evaluation note to every message sent while a paused command exists. The LLM reads the note and decides whether to continue the paused command based on the user's message — no fragile keyword matching for model decisions.

UI icon state uses minimal keyword detection in one place. submit_or_steer_message detects "continue"/"resume" to consume the paused_quarry so the WorkBench icon changes from ⏸ to ▶ and the checkmark ✓ appears on completion. This is a pure UI concern — the model behavior is driven entirely by the LLM note.

Layer	Keyword detection?	Purpose
add_paused_evaluation_note() (centralized)	No	LLM evaluation note for model behavior
dispatch_user_message	No	Uses helper above
steer_user_message	No	Uses helper above
submit_or_steer_message	Yes (1 place)	UI icon state only (▶ and ✓)

What It Does

Frontmatter

• pausable: true in ~/.codewhale/commands/<name>.md enables the pause lifecycle.

Pause / Resume / Cancel

• ESC once pauses a running pausable command. A shared Arc<Mutex<bool>> flag blocks all further tool calls at the engine's pause gate (turn loop).
• Type any message while paused — the message goes through with an evaluation note. The LLM decides whether to continue the paused command. The WorkBench keeps showing ⏸ (via paused_quarry) so the user remembers the paused command.
• "continue"/"resume" — keyword detection in submit_or_steer_message consumes paused_quarry (icon changes to ▶). The LLM evaluation note is still appended so the model understands the intent.
• ESC twice cancels the paused command.
• Cancellation notice injected into non-continue messages sent while paused.

WorkBench lifecycle indicators

• ▶ Play — command is running (yellow)
• ⏳ Pausing — transitioning to paused (tools still draining)
• ⏸ Paused — command is on hold, waiting for resume or cancel
• ✘ Cancelled — command was cancelled by the user
• ✓ Finished — command completed successfully (green, via HuntVerdict::Hunted)

Files Changed

crates/tui/src/commands/user_commands.rs  |  47 +-   pausable frontmatter, state clear
crates/tui/src/core/engine.rs             |  13 +-   shared_paused flag
crates/tui/src/core/engine/handle.rs      |  19 +-   set_paused() -> Arc<Mutex<bool>>
crates/tui/src/core/engine/turn_loop.rs   |  12 +-   pause gate before tool execution
crates/tui/src/core/ops.rs                |   5 +-   SetPaused removed (dead code)
crates/tui/src/tui/app.rs                 |  20 +-   paused, pausable, paused_quarry
crates/tui/src/tui/composer_ui.rs         |  12 +-   EscapeAction::PauseCommand
crates/tui/src/tui/sidebar.rs             | 805 ++++-  WorkBench icons, 80 lifecycle tests
crates/tui/src/tui/ui.rs                  | 226 +++-  dispatch interception, LLM eval note, cancel notice
9 files changed, 1135 insertions(+), 21 deletions(-)

How to Test

Create ~/.codewhale/commands/<name>.md with pausable: true in the frontmatter.

1. Start the command → WorkBench shows ▶ (play, yellow)
2. ESC → WorkBench shows ⏸ (paused), tool calls blocked
3. Type "how are you?" → model responds, does NOT continue paused command, WorkBench keeps ⏸
4. Type "can you please continue the paused command" → WorkBench shows ▶ again, model continues
5. After completion → WorkBench shows ✓ (green checkmark)

To cancel: while paused, press ESC again → WorkBench shows ✘

Test Results

cargo test -p codewhale-tui --bin codewhale-tui pause  -> 18 passed
cargo test -p codewhale-tui --bin codewhale-tui sidebar -> 80 passed (all sidebar tests)

Issue Alignment

Refs #1886, #1887, #1888, #1889, #1891, #1894, #1895, #1917

Builds On

• Phase 1: PR feat: enforce allowed tools for custom slash commands #2326 — custom slash command frontmatter + allowed-tools enforcement
• Phase 2: Hook gate (ToolCallBefore hooks)

---

Paulo Aboim Pinto

[aboimpinto]

All clippy warnings resolved (collapsible_if, unnecessary_map_or, manual_pattern_char_comparison, unnecessary_cast in shell.rs). The git stash snapshot mechanism is acknowledged as a best-effort prototype with known limitations (no restore, sync call). Resume detection was hardened with word-boundary checks (contains " continue "). Manual lock handling in CancelRequest uses direct flag set to avoid Op race. The try_lock for todos/plan clearing is best-effort (logs and continues on failure).

[Hmbown]

this is such a smart update @aboimpinto - wow. thank you for your time and effort!

[aboimpinto]

@Hmbown almost finished ...

[greptile-apps]

aboimpinto has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[greptile-apps]

aboimpinto has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[greptile-apps]

aboimpinto has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[greptile-apps]

aboimpinto has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.

[Hmbown]

Thanks @aboimpinto for the thoughtful lifecycle work here. We landed the safe pausable-command MVP as #2803, merged in 8fe89006700903d1d91868e5e59f8b243f731ec3, with credit preserved in the harvest PR/changelog and a mappable Co-authored-by: aboimpinto <1231687+aboimpinto@users.noreply.github.com> trailer.

The slice that landed keeps pausable: true custom-command frontmatter, the TUI pause/cancel state, and the engine-side pause gate before future tool execution. I intentionally did not carry forward the prototype rollback/stash path because the branch could run git stash push --include-untracked without a matching restore path, which would be risky for contributor/user worktrees.

I am closing this PR as superseded by #2803 so the release train can keep moving, but the broader hook-layer lifecycle intent is still useful future work. Thanks again for pushing this direction forward.