Schutz v0.0.3

The release that made the manager agent tell the truth, and taught Schutz to run commands.

Delegation engine

The manager agent used to claim it had delegated work that never happened. The cause was not the model — it was the app. delegate_task returned a constant success string before the sub-agent had produced a single token, so the model read success and summarised it faithfully. A regex then flagged that honest summary as a lie.

• Delegations now return the sub-agent's actual output. A round starts its delegations first, runs the remaining tools sequentially, then collects the results into the same tool-result batch in the original call order. Delegations complete inside the round, so the round limit is untouched and one tool_use still maps to one tool_result.
• A delegation ledger records every request, rejection, start, and settlement. "Did the manager delegate?" is now a lookup, not a judgement about prose. The old flag was set before the call ran, so unknown-agent, not-connected, and already-busy all counted as "delegated" — the three cases where a user is most likely to be left waiting.
• Nine rejection reasons, localized in Korean, English, German, and Japanese. Each one tells the model what to do instead; a reason without an instruction just gets retried verbatim.
• Delegation timeout (180s): the manager settles honestly and moves on, while the sub-agent keeps working and its proposals still arrive in the review panel.
• Sub-agents now receive context — the files the delegating agent has touched travel with the task, since the delegation prompt is the only channel between them.
• The manager is no longer told it can delegate when no second provider is connected. It used to receive delegation instructions and an empty roster without the tool itself.

Shell commands and dev server preview

• run_command: the agent can run shell commands, with an approval modal in manual mode and live output in the AI log tab.
• Background processes: dev servers keep running independently of the agent that started them, are detected from their output, and open in a preview pane inside the editor. Closing the tab stops the server; quitting the app cleans them up.

Fixes

• Agent replies vanished after tool use. A preview <iframe> firing did-start-loading cancelled the in-flight agent request, and the resulting AbortError was swallowed — indistinguishable from the model saying nothing. Navigation is now gated on the main frame.
• Stop → immediately re-run could make the dying run release the new run's file locks and overwrite its state. Runs are now keyed by run id, and a superseded run skips cleanup entirely.
• Stopping now resolves a pending command approval as a rejection; approvals were not interruptible by abort.
• GPT's subscription path could not edit files — it was missing tool support.
• Tab filenames were unreadable with many files open. Tabs shrank instead of overflowing; the icon, close button, and padding took 61px, leaving 13px for the name. Tabs now keep their size and the strip scrolls, bringing the active tab into view.
• Paper theme left the editor dark — added a light TextMate theme.
• Assorted data-loss and false-success paths: destructive file operations now go through the trash with atomic writes, external modifications are detected, auto-accept no longer reports success it did not achieve, and silent failures surface as toasts.

Chat and motion

• Draft messages survive a restart, ↑ recalls previous messages, a jump-to-latest button appears when you scroll away, and chat opens at the newest message.
• Per-agent chat tabs with per-agent colours, so Claude and GPT no longer blur together.
• Korean/Japanese input no longer sends mid-composition on Enter.
• Language switching, terminal open/close, and project switching animate; the chat no longer shifts horizontally as scrollbars appear.
• AI edits are applied with a typing animation on the real edit path, without remounting the editor.

Branding

• New logo and app icon; the sage brand accent is now separate from semantic green (--ok), so success states no longer fight the brand colour.

Internal

• First test infrastructure in this repository: vitest, with 54 tests covering the engine. The engine is a zero-import pure module type-checked under a stricter config than the rest of the app.