v0.8.45 tracker: control plane and recovery

[Hmbown]

Release theme

v0.8.45 should make agentic terminal work interruptible and recoverable. The user should be able to redirect work mid-flight without corrupting state or losing progress.

This is the control-plane release.

Product thesis

Power is not just automation. Power is being able to stop, steer, inspect, resume, and recover.

In scope

• Consistent cancellation semantics across synchronous tools, shell commands, background jobs, subagents, RLM, and network streams.
• Mid-turn user intervention: queue, pause, redirect, or cancel without pretending the current turn is finished.
• Recovery after crashes or exits: active job state, pending messages, and resumable context.
• Safer shell/process containment on Windows and Unix.
• Clear UI around cancellation pending, cancelled, failed, recovered, and resumable states.
• Control actions that align with the v0.8.43 state model and v0.8.44 task map.

Candidate issue/PR clusters

• pending input/intervention: Pending inputs not consumed until all todos complete in Agent mode — no mid-turn intervention granularity #874
• synchronous tool cancellation: Synchronous tools (file_search, grep_files, list_dir) block turn_loop cancellation #1791, PR fix(grep): respect cancellation token #1839, PR fix(file_search): wrap walker in spawn_blocking + 30s timeout #1790
• shell/process containment: [Windows] Terminal input breaks after exec_shell runs a batch script that spawns long-lived child processes #1690, Windows: shell dispatcher hardcodes cmd.exe, ignoring user's actual shell (PowerShell, pwsh, WSL) #1779, TUI-freeze-Windows-crossterm-poll #1812, [BUG] Windows: Input field completely unresponsive to keystrokes (IME composition event deadlock) #1835
• response-stream pause/data-loss protection: PR fix(installer): pause response stream in all paths to prevent data loss #1860
• background execution and active task wait: v0.8.40: failed shell command can leave TUI feeling stuck with active task_shell_wait/background job state #1737
• VS Code/process crash reports: VS Code crashes or exits unexpectedly when YOLO Agent is running test scripts #1651
• work queue / premature LIVE-state exit reports: Work Queue Sync Lag & Shell PID Hang Causing Premature LIVE-State Exit During Long-Running Tasks #1786

Review lenses

• Security/sandbox reviewer: can cancellation, permissions, process trees, and recovery fail closed instead of pretending success?
• Terminal power user: can they interrupt or redirect without losing focus, input, scroll position, or trust?
• Expert agent operator: can they steer subagents, tools, and long-running work without corrupting the parent turn?
• Systems/tooling builder with taste: are pause, cancel, resume, and recover one coherent control model rather than four special cases?

Acceptance criteria

• Cancelling or redirecting a representative long tool call does not leave the UI in a false live state.
• Recovery after interruption preserves enough state to explain what happened and what can be resumed.
• Windows process/input containment has at least one focused reproduction or documented smoke path.
• The TUI exposes pending/cancelled/failed/recovered/resumable states clearly.
• Regression tests cover the highest-risk cancellation and recovery paths.

[Hmbown]

Updated scope (2026-05-23 integrated roadmap)

v0.8.45 retains the control-plane theme with additions:

• Performance SLOs: cold startup < 500ms, no sync I/O on UI thread, tool dispatch < 100ms p50. Instrument hot paths with tracing.
• Homebrew tap rename: Hmbown/homebrew-deepseek-tui → Hmbown/homebrew-codewhale.
• Codex OAuth provider: merge feat/codex-oauth-provider (modeled on pi-main). Token expiry/refresh/recovery handling.
• Inbox-zero triage: close out backlog items deferred from v0.8.42. Address community PRs from Session 4 harvest.

The existing control-plane scope (interruption, recovery, shell, Windows) stays. See .private/codewhale-integrated-roadmap.md.

[Hmbown]

v0.8.45 control-plane status update

Shipped in #2035 (merged)

The control-plane workstream landed four slices in PR #2035:

1. file_search cancellable — long directory traversals now honour user stop/cancel requests.
2. list_dir cancellable + timeout-safe — blocking worker with a 30-second fallback timeout.
3. Whale-species sub-agent nicknames — stable, human-readable names for sub-agents while preserving raw agent ids.
4. /balance slash command scaffold — registered and dispatchable; still needs the provider capability layer, auth handling, and response parsing before v0.8.45 provider billing: audit cost math and add /balance command #2019 can close.

What remains open

The bulk of #1879 scope is still open:

• Mid-turn intervention (Pending inputs not consumed until all todos complete in Agent mode — no mid-turn intervention granularity #874): queue, pause, redirect, or cancel without pretending the current turn is finished.
• Tool cancellation beyond filesystem: synchronous tools (network, sub-agents, RLM, shell commands) need consistent cancellation semantics (Synchronous tools (file_search, grep_files, list_dir) block turn_loop cancellation #1791, PR fix(grep): respect cancellation token #1839, PR fix(file_search): wrap walker in spawn_blocking + 30s timeout #1790).
• Shell/process containment: Windows & Unix ([Windows] Terminal input breaks after exec_shell runs a batch script that spawns long-lived child processes #1690, Windows: shell dispatcher hardcodes cmd.exe, ignoring user's actual shell (PowerShell, pwsh, WSL) #1779, TUI-freeze-Windows-crossterm-poll #1812, [BUG] Windows: Input field completely unresponsive to keystrokes (IME composition event deadlock) #1835).
• Response-stream pause/data-loss protection: PR fix(installer): pause response stream in all paths to prevent data loss #1860 addressed a narrow slice (checksum manifest race); broader stream protection remains.
• Background execution & active task wait: v0.8.40: failed shell command can leave TUI feeling stuck with active task_shell_wait/background job state #1737.
• Crash recovery: active job state, pending messages, resumable context after exits (VS Code crashes or exits unexpectedly when YOLO Agent is running test scripts #1651).
• UI states: pending/cancelled/failed/recovered/resumable — not yet surfaced clearly in the TUI.
• Performance SLOs: cold startup <500ms, no sync I/O on UI thread, tool dispatch <100ms p50 — not instrumented yet.
• Homebrew tap rename: Hmbown/homebrew-deepseek-tui → Hmbown/homebrew-codewhale — deferred.
• Codex OAuth provider: deferred.
• Regression tests for highest-risk cancellation and recovery paths: none dedicated yet.

Current disposition

Leave open. #2035 is a meaningful first slice (filesystem tool cancellation, sub-agent naming, balance scaffold) but represents ~15–20% of the full control-plane scope. Recommend carrying the remaining items into a future milestone or tracking them as individual issues linked here.

Related v0.8.45 work (not directly #1879)

• v0.8.45: add voice input, RLM session objects, and slash recovery fix #2047: voice input, RLM session objects, slash recovery fix (merged)
• TUI: yield control while waiting on background shell/task work #2009: steer instead of queue when busy-waiting (merged in v0.8.44)

[Hmbown]

v0.8.45 flash release — final disposition

The control-plane work shipped in #2035 represents the filesystem cancellation slice. The broader scope (mid-turn intervention, full tool cancellation, shell containment, crash recovery, performance SLOs, Homebrew rename) is deferred to v0.8.46+.

What shipped in v0.8.45

• feat(file): cancellable list_dir with timeout (control-plane workstream) #2035: cancellable file_search/list_dir, whale-species sub-agent names, /balance scaffold
• v0.8.45: add voice input, RLM session objects, and slash recovery fix #2047: voice input, RLM session objects, slash recovery fix, remember auto-approve fix
• v0.8.45: /model change reverts when picker is dismissed with Esc #2037: model picker Esc applies last-highlighted choice (just closed)
• Goal system Phase 1: create_goal/get_goal/update_goal tools + continuation prompt

Disposition

Leave open. Move to v0.8.46 or keep as the control-plane umbrella tracker for the next release cycle. The 15–20% estimate from the previous update still holds — #2035 was a meaningful first slice but most control-plane scope remains.

[Hmbown]

v0.8.45 final disposition

Closed as won't-complete-in-v0.8.45. The flash release shipped control-plane filesystem tool cancellation (#2035) and several UX improvements. The full control-plane scope (mid-turn intervention, full tool cancellation, shell containment, crash recovery, performance SLOs) is deferred to v0.8.46+.

What landed in v0.8.45

• feat(file): cancellable list_dir with timeout (control-plane workstream) #2035: cancellable file_search/list_dir, whale-species names, /balance scaffold
• v0.8.45: add voice input, RLM session objects, and slash recovery fix #2047: voice input, RLM session objects, slash recovery fix, remember auto-approve
• Goal system Phase 1: create_goal/get_goal/update_goal tools wired to engine
• 6 UX fixes: model picker Esc, sub-agent auto-collapse, shell progress chip, palette refresh, cost persistence
• 3 docs: GUIDE.md, PROVIDERS.md, continuation.md

Tracked for future

• v0.8.46: control plane items, cost/timing polish, low-info rows, provider registry
• v0.8.48: full goal system port (v0.9.0: Port Codex goal system — LLM-as-judge with persistent continuation loop #2058)

Closing this tracker as scope-documented. The remaining control-plane items should be tracked individually in v0.8.46.