v0.8.43 tracker: truth surface

[Hmbown]

Release theme

v0.8.43 should make agent state and evidence legible. The user should never be left staring at Working... without knowing what is running, what is blocked, what last happened, and what basis the agent has for its claims.

This is the cockpit release: expose the machine without making the UI noisy.

Product thesis

The terminal becomes trustworthy when it can answer:

• What is active right now?
• Which tool, command, subagent, request, or background process owns the wait?
• What was the last meaningful event?
• Is this slow, blocked, retrying, cancelled, or waiting for user input?
• What evidence supports the agent's current claim or next action?
• What is the one safest next action or inspection path?

In scope

• A unified runtime state model for active turns, tool calls, shell commands, subagents, RLM sessions, and app-server/serve tasks.
• A compact user-facing state surface in the TUI plus a command or detail pager for deeper inspection.
• Clear stuck-state classification: API stall, process wait, blocked pipe, background job, cancellation pending, approval wait, or UI/input dead zone.
• Evidence records for important claims: file path, command, issue/PR, test, log, memory, or user-provided context.
• Trust labels for verified current, memory-derived, user-reported, unverified, and needs maintainer decision.
• Open log/detail, retry, cancel, and resume affordances wired to the state model where they already exist. Deep interruption semantics can wait for v0.8.45.

Candidate issue clusters to promote from v0.8.42 triage

• stuck/loop reports: TUI read Reading 卡了 #1862, 经常卡住，输入无反应，关闭又会丢失进度 #1830, 项目非常大,问个你好直接卡住 #1827, 始终提示工作中，没有反馈，doctor检查也没问题 #1269, Always stuck, I don't know the status of task! #1190, 卡在工作中，没有回复 #774
• command/job wait problems: v0.8.40: failed shell command can leave TUI feeling stuck with active task_shell_wait/background job state #1737, Work Queue Sync Lag & Shell PID Hang Causing Premature LIVE-State Exit During Long-Running Tasks #1786, Deadlock: process hangs unrecoverably when API connection stalls — blocked on pipe_read with closed write end #1472
• subagent/RLM visibility gaps: Subject: Sub-agent 120s API timeout renders agent_open nearly unusable (v0.8.39) #1806, sub-agent deep-analysis results are lossy: parent only gets compressed summary, cannot retrieve full child output after completion #1738
• tool cancellation: Synchronous tools (file_search, grep_files, list_dir) block turn_loop cancellation #1791, fix(grep): respect cancellation token #1839, fix(file_search): wrap walker in spawn_blocking + 30s timeout #1790
• loop guard status correctness: PR fix(engine): report loop guard halt as Failed instead of Completed #1859
• issue/PR close-out and source verification from v0.8.42 tracker: inbox-zero triage and backlog close-out #1876

Review lenses

• Terminal power user: can they tell what owns the wait without losing keyboard flow?
• Expert agent operator: can they distinguish a slow model, blocked tool, stalled stream, running subagent, and false completion?
• Security/sandbox reviewer: does the state surface avoid overstating trust or hiding tool execution?
• Brave beginner: does the UI say what happened and what to do next without requiring internal vocabulary?

Acceptance criteria

• The TUI can show the current active unit of work and last event without reading raw logs.
• A stalled turn has a classified reason or an explicit unknown state with diagnostic pointers.
• Failed, cancelled, timed-out, and completed states are visually and semantically distinct.
• At least three historical stuck reports have either a verified fix, a canonical duplicate, or a documented cannot-reproduce disposition.
• Final task summaries distinguish verified results from unverified assumptions.
• Focused tests cover the state/evidence transitions touched by the release.

[Hmbown]

Updated scope (2026-05-23 integrated roadmap)

v0.8.43 absorbs the registry publishes and repo rename that were originally in v0.8.42:

• Publish 14 codewhale-* crates to crates.io (first-time publishes)
• Publish codewhale@0.8.43 to npm with binary version pinned
• Publish deepseek-tui@0.8.43 deprecation shim
• Rename GitHub repo Hmbown/DeepSeek-TUI → Hmbown/codewhale
• Merge web rebrand PR (R7) and docs domain flip (R8)
• Verification: npm view codewhale@0.8.43, cargo search codewhale-cli, gh repo view Hmbown/codewhale

The original "truth surface" work moves to v0.8.45 (performance track) and v0.8.46 (verifier preview).

See .private/codewhale-integrated-roadmap.md for the full plan.

[Hmbown]

Added #1972 to the v0.8.43 truth-surface scope: resizable thinking blocks, visible preview counts, and a concrete implementation prompt for configurable reasoning preview controls. This is intentionally scoped as an interaction/config primitive rather than a broad settings-panel redesign.

[Hmbown]

Added #1975 to v0.8.43 truth-surface scope: make the Tasks sidebar inspectable, expandable, and copyable. It captures the Tasks panel problems from current smoke testing: truncated IDs, ambiguous active/running counts, clipped tool output without recovery, copy-unfriendly recent-tool rows, and missing yank/detail affordances.

[Hmbown]

v0.8.43 scope update (2026-05-24)

Two notable additions to the truth-surface theme since this tracker opened, plus a redistribution that drops v0.8.43's open count from 60 to 27 so the milestone fits an actually shippable window. See the milestone body for the full updated In-scope / Out-of-scope / Definition of done.

New work surfaced under this tracker

• v0.8.43: make Tasks sidebar inspectable, expandable, and copyable #1975 — Tasks sidebar inspectability. Make the Tasks panel a workbench surface (full recoverable IDs, replace ambiguous active / running wording, structured two-line recent-tool rows, expand/detail paths, y / Y yank, copy-paste-friendly rows, terminal-native throughout). Discovered from v0.8.42 smoke testing.
• v0.8.43: Goal mode — persistent objective/workflow surface #1976 — Goal mode (persistent objective surface) + Fin wakeup/verifier. A new product lane: replace the narrow /goal command with a persistent objective surface where the agent can clarify, plan, and keep working until done, blocked, or explicitly stopped. Fin (DeepSeek V4 Flash, thinking off) acts as a cheap binary-perspective wakeup/verifier — does the work pass tests, build, docs, install state, release gates? Goal mode is orthogonal to Plan / Agent / YOLO (the three execution modes) and to --model auto (model routing). Fin wakeup here is the Goal-mode-specific case; the general verifier-native completion preview lives in v0.8.46.

Redistribution from the v0.8.41 absorb (60 → 27 open)

• → v0.8.45 (control plane + provider parity + perf + Homebrew): hostability / cancellation / Windows / long-session cluster ([BUG] Windows: Input field completely unresponsive to keystrokes (IME composition event deadlock) #1835, v0.8.40: failed shell command can leave TUI feeling stuck with active task_shell_wait/background job state #1737, [Windows] Terminal input breaks after exec_shell runs a batch script that spawns long-lived child processes #1690, SSE多智能体并行在windows11下依旧45s超时，并且还出现了UI错乱的问题 #1679, VS Code crashes or exits unexpectedly when YOLO Agent is running test scripts #1651, 使用windows wsl 安装后运行报错 #1596, Bump portable-pty to 0.9 #1531, Synchronous tools (file_search, grep_files, list_dir) block turn_loop cancellation #1791, Subject: Sub-agent 120s API timeout renders agent_open nearly unusable (v0.8.39) #1806, Deadlock: process hangs unrecoverably when API connection stalls — blocked on pipe_read with closed write end #1472, Work Queue Sync Lag & Shell PID Hang Causing Premature LIVE-State Exit During Long-Running Tasks #1786, 执行大文本处理工程后会话中断卡死 #1425, 项目非常大,问个你好直接卡住 #1827, 合并分析报告保存文档巨慢 #1732, [Bug] Conversation cannot be triggered on Windows via npm global install (Stuck at "Working") #765) plus PRs fix(grep): respect cancellation token #1839, fix(tools): replace cross-await RwLock with Semaphore to prevent deadlock #1856.
• → v0.8.46 (tool studio + verifier preview): copy/select, balance/cost/cache display, terminal preview, sub-agent UX, model-loop (TUI: terminal-native copy includes visual line breaks from wrapped output #1853, Feature: Display API account balance #1735, 建议token成本估算新增加几个货币单位 #1607, 支持底部状态栏自定义配置（[statusline]），显示 API 余额等实时信息 #1551, 输入缓存命中率太低了 #1177, Git-bash/window terminal preview Multi-Line #1687, 执行结果输出和思考预览改善 #1682, sub-agent deep-analysis results are lossy: parent only gets compressed summary, cannot retrieve full child output after completion #1738, 模型出现了循环问答幻觉 #1863).
• → v0.8.47 (continuity + provider abstractions): project-context (feat: add DEEPSEEK.md as project context for harness integrations like superpowers #1851), auto-compact threshold (feat: configurable auto-compact threshold with Ctrl+L keybinding #1722), token-consumption context regression (token消耗增大了很多 #743), plus PR feat: add DEEPSEEK.md as project context file #1852.
• → v0.8.48 (presence + shim-drop): macOS title layout ([UI Bug] macOS 终端标题悬浮居中，布局未贴顶 #1834) plus PRs fix: 修复 macOS 终端标题悬浮居中，改为贴顶布局 #1837, feat(tui): add "Terminal" theme #1831.

v0.8.43 keeps: truth-surface anchors and refactor enablers (#1972, #1894, #1897, #1887, #1886, #1884, #1961), the generic stuck-state bugs that need stall-reason fixes (#1190, #1269, #1862, #1147, #1830, #861, #1435, #689, #1439), release / install hygiene (#1869, #1858), plus PRs that fit truth-surface or release hygiene (#1859, #1836, #1860, #1870, #1844, #1850, #1875).

Anti-scope confirmed kept: ~/.deepseek/, DEEPSEEK_* env vars, api.deepseek.com, model IDs, ProviderKind::Deepseek, CNB mirror namespace, historical CHANGELOG entries. /model auto stays as model routing — not an execution mode.

[Hmbown]

v0.8.43 released. GitHub Release: https://github.com/Hmbown/CodeWhale/releases/tag/v0.8.43. Goal mode, receipts, stall reasons, decision cards, tasks sidebar inspectability, thinking toggle, sub-agent drain, grep_files cancellation, npm installer fix, Ctrl+Z restore, Wayland clipboard, README contributor refresh. npm + crates.io publishes pending manual 2FA.