Skip to main content

Release 0.4.9 — HITL Tool Approval, GridFS, and Docs Redesign

Sri Aradhyula
Sri Aradhyula

Released: 2026-05-07 Chart: oci://ghcr.io/cnoe-io/charts/ai-platform-engineering:0.4.9 Previous release: 0.4.8

Highlights​

0.4.9 ships Human-in-the-Loop tool approval for dynamic agents, a GridFS-backed store for large tool results, custom agent color themes, an improved agent-selection UX, an automated release-docs coding-agent skill, and a fully redesigned documentation site.

What's New​

Human-in-the-Loop (HITL) Tool Approval​

  • Configurable interrupt_on per agent — admins can require explicit approval before specific MCP tools execute; the agent pauses mid-run and waits for a human accept/reject
  • Tool approval card UI — new ToolApprovalCard component surfaces the pending tool call with its arguments; timeline expands during wait and collapses on resume
  • Interrupt config picker — new "Advanced" step in the agent editor lets you select which tools require approval without touching YAML
  • Invoke endpoint hardening — /invoke now detects HITL interrupts and returns a clear error rather than hanging (#1351 by @subbaksh)

GridFS Backend for Large Tool Results​

  • MongoDBGridFSStore — stores tool results in MongoDB GridFS, bypassing the 16 MB document limit that caused checkpoint failures for large LLM outputs
  • Enabled by default via USE_GRIDFS_BACKEND=true; can be toggled off for smaller deployments
  • format_file builtin tool — reformats single-line blobs into newline-delimited output, making grep-friendly file searching possible from within an agent run
  • TTL index — files auto-expire after 48 hours; no manual cleanup needed (#1351 by @subbaksh)

Custom Agent Color Themes​

  • Custom theme option — replaces the fixed "midnight" preset; each agent can now carry its own gradient and bot-icon tint
  • Color picker popover — live preview of from / to gradient colors and accent; AI suggest can generate a full color scheme from a natural-language description
  • Avatar rendering updated everywhere to pick up per-agent accent color (#1351 by @subbaksh)

Agent Selection Search & Scroll​

  • Search input — auto-focuses when the agent dropdown opens; filters both the Platform Engineer default and all dynamic agents by name or description
  • Scrollable list — agent list is capped at max-h-64 with overflow-y-auto; no more overflow when many agents are configured
  • Search state clears on selection or close (#1361 by @kevkantes)

release-docs Coding-Agent Skill​

  • New .claude/skills/release-docs/SKILL.md — invoke /release-docs in Claude Code to generate a release blog post and upgrade guide from git history; wired into docs-release.yml CI so a post is opened automatically on every stable tag (#1356 by @sriaradhyula)

Documentation Site Redesign​

  • 3-column hero — tagline + CTAs | live demo GIF with fullscreen lightbox | install code blocks (curl / Kubernetes Helm / Kind)
  • Live GitHub star count with rate-limit guard
  • Compact 11-feature grid replacing the previous 6-tile layout
  • Multi-version docs — 0.4.9 as Latest, main 🚧 accessible at /docs/next/
  • Repo Ops section under Development — CI workflows reference, release guide, and coding-agent skills catalogue
  • Prebuild flow doc — full guide to prebuild/* image and Helm chart CI
  • GIF demo hosted on GitHub release assets (not a binary blob in the repo) (#1359 by @sriaradhyula)

Bug Fixes​

  • ingestor: prevent UnboundLocalError in the finally block when client.initialize() fails before the client variable is set — caused silent crashes on misconfigured ingestor startup (#1357 by @sriaradhyula)
  • docs: escape MDX angle bracket in spec plan template; add docs-build-check.yml workflow to catch broken Docusaurus builds on PRs touching docs/ (#1360 by @sriaradhyula)
  • dynamic-agents: centralised LLM instantiation via llm_clients.get_llm — fixes a subagent model config bug where the wrong model was used when AGENT_MODEL differed from SUPERVISOR_MODEL; inlined StreamingMixin; removed dead models and state (#1351 by @subbaksh)

Breaking Changes​

No breaking changes. Drop-in upgrade from 0.4.8.

Known Issues​

None known at this time.

Upgrade​

helm upgrade ai-platform-engineering \
  oci://ghcr.io/cnoe-io/charts/ai-platform-engineering \
  --version 0.4.9 \
  -f your-values.yaml

Upgrade Guide: 0.4.8 → 0.4.9​

Overview​

0.4.9 adds HITL tool approval and GridFS backend for dynamic agents, plus UI improvements and docs updates. No Helm values changes are required — new features default to enabled and existing deployments continue to work without modification.

Helm Values Changes​

No breaking Helm values changes. Drop-in upgrade — no values.yaml edits required.

New Optional: GridFS Backend​

GridFS is enabled by default. To disable it (e.g., if your MongoDB instance doesn't support GridFS):

dynamic-agents:
  config:
    USE_GRIDFS_BACKEND: "false"

New Optional: HITL Tool Approval​

To require human approval before a specific tool executes, set interrupt_on in the agent config via the agent editor UI, or in your agent YAML:

interrupt_on:
  - tool_name: "my-sensitive-tool"

No Helm chart changes needed — this is per-agent configuration managed at runtime.

Multi-Version Docs​

The documentation site now serves 0.4.9 as the default (Latest) version. The previous main-based docs are accessible at /docs/next/. If you have bookmarks to versioned doc URLs, they will continue to work.