Repository navigation
Kintsugi v0.4.1
Kintsugi v0.4.1
v0.4.0 was not published; this is the first published release of the 0.4 minor series.
Code mode templates and one-shot agents
Prompt templates in .kintsugi/prompts or .pi/prompts can declare codemode: true. Their JavaScript runs only when the slash command is explicitly submitted, with args and argumentsText available. Saving a template starts no work. Code mode templates do not accept attachments.
Scripts can call tools.agent({ prompt, schema }) for a fresh, task-owned conversation and a validated JSON answer. Calls retain their conversation and deduplicated inputs across recovery; Stop ends their owned work. Their transcripts open from the script's call details, and usage contributes to the parent Agent's total.
A read-only transcript does not mean side-effect-free tools. One-shot agents can use a selected subset of read, write, edit and bash, intersected with their caller's enabled capabilities. They cannot delegate, run further scripts, access another Agent's plan or obtain undispatched Review comments. Existing workspace credential and network restrictions still apply.
Admission is limited to eight concurrent one-shot agents per workspace. Each call has a five-minute deadline, three attempts to produce a valid answer, and a 32k output-token threshold checked between provider requests; a single request may cross that threshold. Schemas are limited to 32k characters, 1,024 nodes and depth 32, without regex patterns or references. Answers are limited to 64k characters, 4,096 nodes and depth 32. Code mode retains its two-worker host limit, counted through actual worker exit, and 256 MiB VM limit per worker.
Unified Background
Background groups delegated work and Claude Code tasks above the conversation. Commands and monitors expose output in place; subagents and workflows open their own pages, with progress, transcripts and supported Stop controls. Workflow phases include running and queued agents, and link to started agents' transcripts.
One-shot agents belong to their script rather than the Background list. Existing Subagents routes and subscriptions remain available. Newly exposed Claude background journal previews require regular, non-symlinked files and safe task IDs; they retain at most the last 2 MiB, with 16 cached tails. Source journals are not changed.
Live interaction and usage
- Live collection updates send changed rows without resetting unchanged permanent rows.
- Transcript following resumes when the reader actually moves toward the end. Opening a disclosure or expanding thinking pauses following; live updates leave paused reading positions alone.
- Composer tap highlighting is suppressed on iOS.
- Anthropic usage cards show extra-usage status and available allowance details.
- Browser-local idle timestamps remain localized after Turbo morphs, while retaining their UTC instant. Background elapsed-time facts have an immediate server-rendered value.
Verification and operation
Verification includes frozen installation, lint/types, asset builds, the full test suite and native-Harness lifecycle checks. Manual acceptance uses a local faux provider and a fake Claude transport, not production inference. Composer layout acceptance passed 89/89 checks in six explicitly partial desktop/mobile Chrome scenarios. Production-provider inference and physical/iOS soft keyboards are not live-qualified.
System capability remains 3, app minimum 0, network-policy reader 2 and resource-policy reader 1. Pi remains pinned to 1.3.0. Provider routing/caching defaults, explicit Review delivery, credential grants, custom sources, version pins and paired rollback are unchanged. No application environment variable is added.
Publication advances both latest and stable after native AMD64/ARM64 verification. No existing installation is updated or restarted. Before downgrading, stop active scripts and one-shot agents and back up app data; image rollback does not restore journals or settings.
These eight upstream changes were adapted from Lucas Meijer's AgentsInTheCloud. Existing Lucas Meijer and Larry Hudson attribution and MIT notices remain intact.