This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+2 more
Affected surfaces
Summary
AI summaryThe dev board now spawns a headless Claude Code agent that streams output live when a gate is approved or Run is pressed.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Security | Medium |
Enforces same-origin guardrails for agent execution. Enforces same-origin guardrails for agent execution. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Security | Medium |
Restricts project location to $HOME directory. Restricts project location to $HOME directory. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Security | Medium |
Sets default permission `acceptEdits` and requires opt-in flag for full autonomy. Sets default permission `acceptEdits` and requires opt-in flag for full autonomy. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Feature | Medium |
Adds headless Claude Code agent spawning and live streaming on dev board approval or Run button. Adds headless Claude Code agent spawning and live streaming on dev board approval or Run button. Source: llm_adapter@2026-06-10 Confidence: high |
— |
| Feature | Low |
Adds configurable binary path for the Claude Code agent. Adds configurable binary path for the Claude Code agent. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Performance | Low |
Limits agent runs to one per project, returning 409 on conflicts. Limits agent runs to one per project, returning 409 on conflicts. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Performance | Low |
Imposes hard timeout with SIGTERM then SIGKILL for agents. Imposes hard timeout with SIGTERM then SIGKILL for agents. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
| Performance | Low |
Caps output buffer to a 2000-line ring buffer. Caps output buffer to a 2000-line ring buffer. Source: granite4.1:30b@2026-06-10-audit Confidence: low |
— |
Full changelog
The dev board becomes a pult, not just a mirror: approving a gate (or pressing Run) spawns a Claude Code agent headlessly in the project and streams its output live into the board.
- Spawn + stream —
claude -p "<prompt>" --output-format stream-jsonin the project cwd; events parsed (assistant text · tool_use · result) and pushed asevent: agentover SSE. Endpoints/api/agent/run|stop|status; gate-approve takes optional{ runAgent, agentPrompt }. - Guardrails — same-origin only · project must live in
$HOME· one run per project (409) · hard timeout (SIGTERM→SIGKILL) · 2000-line ring buffer · configurable bin. Permission defaultacceptEdits; full autonomy opt-in viaGREAT_CTO_AGENT_DANGEROUS=1. - UI — a Run agent panel (prompt · Run · Stop · live stream) + an Approve + ▶run button on gate cards.
Verified with a stub bin (streaming + all four guardrails + Stop) and a real claude run. Lib 348/348.
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
About great_cto
Engineering-management layer of 34 specialist AI agents covering the full SDLC (architect, PM, senior-dev, reviewer, QA, security, devops, L3-support + 18 archetype-specific reviewers) with auto-detected archetypes and compliance gates (PCI-DSS, HIPAA, FedRAMP, GDPR, EU AI Act). Runs in Claude Code, Cursor, Codex CLI, Aider, and Continue via AGENTS.md + MCP. MIT.
Related context
Related tools
Earlier breaking changes
- v2.72.2 Removes Autopilot sidebar item and Operate topbar button from dev board.
- v2.55.0 Hard-gates the Build board; redirects invite sessions to Operate for operators.
- v2.43.0 Runtime now blocks autonomous execution of irreversible actions without prior human checkpoint.
- v2.43.0 Adds reversible and blastRadius fields to every flow step.
- v2.32.0 Removed AgentShield scanner and its CLI commands.
Beta — feedback welcome: [email protected]