Idea-to-build

v0.1.3 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

Published 5d AI Agents & Assistants

✓ No known CVEs patched

✓ No known CVEs patched in this version

Topics

ai-agents brainstorming claude claude-code cowork-plugin mcp

+3 more

methodology multi-agent prompt-engineering

Summary

AI summary

Sub-agent dispatch now correctly targets the real agent-launch tool.

Type	Severity	Summary	CVE
Feature	Low	Added `examples/food-delivery-comparison.md` showcasing a trace-verified isolated run. Added `examples/food-delivery-comparison.md` showcasing a trace-verified isolated run. Source: granite4.1:30b@2026-06-03-audit Confidence: low	—
Bugfix	Medium	Sub-agent dispatch now targets the correct agent-launch tool, fixing incorrect targeting to Task tools. Sub-agent dispatch now targets the correct agent-launch tool, fixing incorrect targeting to Task tools. Source: llm_adapter@2026-06-03 Confidence: low	—
Bugfix	Medium	Sub-agent dispatch now correctly targets the real agent-launch tool, avoiding collision with to-do tools. Sub-agent dispatch now correctly targets the real agent-launch tool, avoiding collision with to-do tools. Source: granite4.1:30b@2026-06-03-audit Confidence: low	—
Refactor	Low	Re-recorded `examples/hyperliquid-wallet-comparison.md` from a trace-verified isolated run. Re-recorded `examples/hyperliquid-wallet-comparison.md` from a trace-verified isolated run. Source: granite4.1:30b@2026-06-03-audit Confidence: low	—

Full changelog

The important release: the multi-agent isolation that the whole methodology rests on now actually fires, and the deep examples prove it.

Sub-agent dispatch now targets the real agent-launch tool. The protocol said "the Task tool," which in Cowork collides with the TaskCreate/TaskUpdate to-do tools — so the coordinator was tracking a checklist instead of spawning an isolated critic, and Phases 5/6 ran with no real context isolation. It now names the correct tool per host (Agent in Cowork, Task in Claude Code), forbids the to-do tools, and requires a verification step that a separate agent actually ran.

examples/food-delivery-comparison.md — the fullest "can it say no?" arc, and the first trace-verified isolated run: the cold critic returns NO-GO on the head-on idea, then GO-with-conditions on the pivot.

Re-recorded examples/hyperliquid-wallet-comparison.md from a trace-verified isolated run (Phase 5 shows ran an agent) that returns NO-GO. The earlier version documented a run whose dispatch had used the to-do tool, so its "isolated" claim wasn't accurate; now it is.

Both deep examples (Hyperliquid, food delivery) are now genuinely isolated and trace-verified.

Full changelog: CHANGELOG.md

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track Idea-to-build

Get notified when new releases ship.

About Idea-to-build