feat(tui): run goals in Terminal Studio

[Delta456]

Bare appclaw and appclaw "a goal" now open Terminal Studio in a new goal mode, instead of the readline prompt plus single-screen Ink run. Same shell as --tui, same device session, same palette and history — only what a plain (non-slash) line means differs: an instruction to record, or a goal to run.

That change pulled in four others: the goal pipeline had to come out of index.ts before two callers could share it; the prompt had to stop eating keyboard shortcuts before goal mode could have any; a finished run had to leave its summary somewhere reachable; and a bad .env had to stop killing the process before any of it could be reported on screen.

---

Goal mode

store.mode is 'record' | 'goal'. --tui opens in record (a plain line is one deterministic instruction, appended to the recording); bare appclaw opens in goal (a plain line is a goal, run through the full planner and drawn on a new run screen beside the device stream). /mode switches without dropping the session.

• packages/cli/src/goal-session.ts (new) — decomposeGoal → per-sub-goal orchestration → runAgent → journey summary. MCP, LLM and AppResolver are dependencies rather than things it creates, so the one-shot CLI and the resident shell drive the same pipeline. It deliberately does not own device setup, Ink mounting, --export writing or process exit — those differ between a one-shot run and a resident shell. index.ts is 429 lines lighter.
• attachRunRenderer() — two Ink apps cannot share stdout, so goal mode never calls activateInk(); it borrows only the renderer seam and mounts RunScreen as one of its own screens.
• AgentOptions.signal — cooperative cancellation, checked at each step boundary. A stop lands after the action in flight finishes, which is why the screen says "stopping…" rather than implying the keypress was instant.
• appclaw "goal" holds the finished screen until a key is pressed before exiting 0/1. Leaving the alternate screen erases the summary and the last device frame, which is the whole point of having watched. APPCLAW_TUI=off keeps the unattended path.

The palette is hidden in goal mode, not filtered

A plain line there is a goal, so a standing command list is a menu for nothing — the transcript gets its rows instead. The column's total height is unchanged, which is what keeps the stream panel where the frame loop expects it. The list reappears as soon as the line starts with /, where it is a live filter, and a wrapped prompt then borrows its extra rows from the transcript. layout.ts owns those numbers (paletteRows, transcriptRows, inputLineBudget, inputLineCount) so MainScreen and CommandPalette cannot disagree about how tall the prompt is.

Stream shortcuts, and the input rewrite they needed

^r starts or resumes the stream, ^p pauses/resumes, ^x closes it — bound on both the goal prompt and the run screen by tui/stream-keys.ts, which holds the chord table, the handler and the status-bar hints together so a binding cannot outlive its advertisement. Everything the stream says about itself now names a chord rather than a slash command, because goal mode hides the palette.

They are ctrl chords because the goal prompt is a focused text input, where a bare p is part of the goal being typed. That is why components/PromptInput.tsx replaces ink-text-input: that component inserts any ctrl chord it does not recognise as a plain letter (^p types p), so no screen holding a focused prompt could own a shortcut. PromptInput ignores every ctrl/meta chord (ctrl+c included, so TuiApp still quits), leaves arrows/tab/page keys to the screen, and snaps the cursor to the end when the value is replaced from outside — which also fixes history recall and tab completion landing the next character mid-string.

StreamStatus gained paused, genuinely distinct from idle: closing stops the loop, removes the temp dir and deletes the transmitted kitty image, whereas pausing only gates capture and leaves the frame on screen for a free resume.

/export and the journey summary survive a run

The run screen was the only place the journey summary was drawn, so leaving it took the sub-goal breakdown, tokens and cost with it — re-running the agent was the only way to see any of it again. summariseOutcome() (tui/goal-summary.ts) writes them into the transcript as the result entry's detail block, sub-goal results included, since what the agent found is the reason the run happened.

/export now works in both modes, from different material: the recorded step list in record mode, the last goal run in goal mode, through the same writer appclaw "goal" --export uses. The resident shell is where you iterate on a goal until it does the right thing, which is exactly when you want it frozen into a deterministic spec. The run's platform is captured with the outcome rather than read at export time, so a /device switch in between cannot write an iOS spec for an Android run.

A rejected .env value no longer crashes with a stack trace

Config is built at import time, so a throw there escaped every try/catch in the CLI and Node printed a raw ZodError and a stack trace because someone typed LLM_PROVIDER=claude.

safeLoadConfig() reports instead of throwing, dropping only the offending keys so their schema defaults apply and the rest of the environment survives — enough to say what is wrong, never enough to run on. loadConfig() still throws for the SDK, now with a readable message.

Who reports it is decided in one place, opensTerminalStudio(cliArgs): the shell can rewrite .env, so it gets the problem on SetupScreen and can fix it in place; every other path takes printError + exit 1 rather than running on defaults the user did not choose. Settings shows a rejected key's raw .env value rather than the default it fell back to (the typo is the thing being corrected) and appends any offending key not already listed, so "press enter to edit it here" is never a dead end.

$ LLM_PROVIDER=claude AGENT_MODE=hybrid APPCLAW_TUI=off appclaw --flow x.yaml
  ✗ Invalid configuration
    LLM_PROVIDER: Invalid enum value. Expected 'anthropic' | 'openai' | 'gemini' | 'groq' | 'ollama', received 'claude'
    AGENT_MODE: Invalid enum value. Expected 'dom' | 'vision', received 'hybrid'

Smaller fixes

• StepLine cells got flexShrink={0} + wrap="truncate" — in a narrow column Yoga reflowed the fixed-width grid onto a second row ([1/3 over 0]), scrambling the row and costing the caller a row its height budget never reserved.
• RunScreen sizes from RunWidthContext rather than stdout.columns, which are the same thing only when the run owns the whole terminal; in a pane the summary box drew wider than its column and overflow: hidden sliced its border off.
• StatusBar breadcrumb no longer clipped to "Goa" at 100–118 columns.
• Wordmark extracted from WelcomeScreen.

Docs

README.md, CLAUDE.md, .agents/skills/use-appclaw-cli/SKILL.md and landing/public/usage.html updated for goal mode, the chords, /export and the config behaviour.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_0163wRmYBwSsV8xUDmuiEZio

[github-actions]

🎉 This PR is included in version 2.6.0 🎉

The release is available on:

• v2.6.0
• GitHub release

Your semantic-release bot 📦🚀