SuperAPI AI Directory

env-capability-discovery

Agents & models

View repository

Updated 10-15 21:47

About this skill

Environment Capability Discovery

Internal shared skill. A workflow runs on heterogeneous platforms (Claude Code, OpenCode, Cursor, …) with different installed skills/agents. This skill defines how to discover and invoke optional enhancements instead of relying only on the workflow's own built-in flow.

Reference semantics: weak reference by default — referencing workflows normally do NOT list this skill in frontmatter dependencies; when it is unavailable, they skip discovery silently and run their original flow — no error, no abort. A workflow MAY explicitly declare it as a strong dependency (as solve-workflow does); that workflow's prerequisite check then guarantees availability, and a missing skill aborts the workflow at startup.

When to scan

Once, at workflow startup (before the first working stage). Record the result for the rest of the session (the referencing workflow defines where — conversation context or a state file); never rescan.

How to scan

Use the current platform's skill-discovery capability to list available skills/agents/plugins:

  • Check the available-items listing in the system prompt, or use the platform's skill-listing tool.
  • Graceful degradation: if no discovery method exists, skip enhancement discovery and run the original flow.

Capability types and matching keywords

Match against each candidate's name and description:

Capability Keywords
🔍 Debugging analysis debug, root-cause, investigate, systematic-debugging, analyze problem
🌐 Web research research, web, look up, investigate, web-research, effective-web-research
💡 Solution design brainstorm, design, architect, explore solution
📋 Code review code-review, review, requesting-review
📝 Planning plan, writing-plan, make plan
⚡ Execution orchestration execute, executing-plan, subagent, parallel
🧪 Test-driven test, tdd, test-driven
🔧 Build repair build-fix, build, linter, type-check
✅ Completion verification verify, verification, complete
🖥️ Browser control browser, ego-browser, playwright-mcp, agent-browser, chrome-devtools, cdp
🌿 Branch management worktree, branch, git-worktree

This table contains no stage numbers: each referencing workflow keeps its own small capability→stage mapping table (stage numbering differs per workflow) alongside its reference note.

Handling scan results

  • Matched: record it; load and invoke at the workflow's mapped points.
  • No match: skip silently; run the original flow. Never error, never block.
  • Multiple matches of one type: prefer the one with the more specific description.

Invocation principles

  1. Progressive enhancement, never a replacement: enhancements assist; they do not change stage order or gate logic.
  2. Read-only stages stay read-only: an enhancement never grants Edit/Write to a stage that forbids it.
  3. Enhancement failure never blocks: on error, log a warning and continue the original flow.
  4. Read the skill's current doc before invoking: never invoke from memory — rules may have changed.
  5. frontmatter dependencies are out of scope: strong dependencies are guaranteed by the workflow's prerequisite check; do not rediscover them here.