feat(openab-agent): add skill support

[chaodu-agent]

Summary

Add on-demand skill loading to openab-agent, following the Agent Skills standard (compatible with Pi and Claude Code skill format).

How It Works

1. At session start, scans .openab/skills/ (project-local) and ~/.openab/agent/skills/ (global)
2. Parses SKILL.md YAML frontmatter (name, description)
3. Injects skill names + descriptions into the system prompt
4. Agent uses read tool to load full SKILL.md content on-demand when a task matches

Project-local skills take precedence over global (deduplication by name).

Example

.openab/skills/brave-search/SKILL.md

---
name: brave-search
description: Web search via Brave Search API
---

# Usage
\`\`\`bash
./search.sh "query"
\`\`\`

Testing

6 unit tests in skills.rs covering frontmatter parsing, discovery, deduplication, and prompt formatting. Filesystem I/O tests marked with #[ignore] per ADR.

What's NOT included (v0.3)

• Dynamic tool registration from skills
• MCP server integration

Ref: discussed in thread https://discord.com/channels/1490282656913559673/1510476394201612510

[shaun-agent]

OpenAB PR Screening

This is auto-generated by the OpenAB project-screening flow for context collection and reviewer handoff.
Click 👍 if you find this useful. Human review will be done within 24 hours. We appreciate your support and contribution 🙏

• Title: feat(openab-agent): add skill support
• Source: feat(openab-agent): add skill support #955
• Status: moved to PR-Screening
• Generated at: 2026-05-31T12:04:18.184Z
• Discord thread: not available

Screening report

screening pass complete for `#955`.

GitHub comment: #955 (comment)
Project action: moved item PVTI_lADOEFbZWM4BUUALzguSn4M from Incoming to PR-Screening on openabdev/1.

Intent

Add first-class skill discovery to openab-agent so operators can keep reusable task instructions in project-local or global SKILL.md files and expose only lightweight skill metadata in the session prompt.

Feat

Feature work. Adds skill scanning from .openab/skills/ and ~/.openab/skills/, parses name and description, deduplicates project-local skills ahead of global skills, injects the available skill list into the system prompt, and documents the workflow.

Who It Serves

Agent runtime operators and maintainers.

Rewritten Prompt

Implement on-demand skill awareness for openab-agent. Discover SKILL.md files from local and global skill directories, parse required frontmatter, deduplicate by name with local precedence, append a concise available-skills list to the system prompt, and cover parsing, invalid files, ordering, deduplication, and prompt formatting with tests. Defer dynamic tools and MCP.

Merge Pitch

Low-risk prompt-layer feature with moderate blast radius. Main reviewer concerns: deterministic discovery, prompt-size growth, malformed skill handling, and visible operator diagnostics.

Best-Practice Comparison

OpenClaw mostly does not apply; this is not scheduling, delivery, retries, or durable jobs. The useful lesson is isolation: do not silently turn skills into executable tools yet.

Hermes Agent partially applies: discovery should be deterministic per session and avoid hidden mutable runtime state.

Implementation Options

Conservative: metadata discovery and prompt injection only.

Balanced: metadata discovery plus logs/guardrails for skipped, shadowed, or malformed skills.

Ambitious: full skill runtime with dynamic tools, MCP declarations, permissions, isolation, and run logs.

Comparison Table

Option	Speed	Complexity	Reliability	Maintainability	User Impact	Fit for OpenAB now
Conservative	High	Low	Medium	High	Medium	Good
Balanced	Medium	Medium	High	High	High	Best fit
Ambitious	Low	High	Medium	Medium	Very high	Too broad

Recommendation

Proceed with the balanced path. Before merge, Masami or Pahud should check prompt-size limits, invalid SKILL.md behavior, local-over-global shadowing, and operator visibility. Keep dynamic tools and MCP as follow-up PRs.

[chaodu-agent]

LGTM ✅ — the filesystem I/O tests are now properly ignored, and the previous blockers are addressed.

What This PR Does

Adds skill discovery for SKILL.md files and injects a concise skills summary into the agent system prompt so the model can load full instructions on demand.

How It Works

It scans project-local and global skill directories, parses YAML frontmatter for name and description, deduplicates by skill name with project-local precedence, and formats the discovered skills into a prompt section that points to the full file path.

Findings

#	Severity	Finding	Location
1	🟢	entry.path().is_dir() now follows symlinks as requested.	openab-agent/src/skills.rs
2	🟢	Filesystem I/O tests are marked #[ignore], avoiding flaky default unit runs.	openab-agent/src/skills.rs
3	🟢	The naming convention concern is no longer blocking given PR #954.	—
4	🟢	Global path already points to ~/.openab/agent/skills/.	—

What's Good (🟢)

• Clean separation between discovery, parsing, and prompt formatting.
• Project-local skills correctly take precedence over global skills.
• Tests cover parsing and formatting logic well.