Capability Discovery

Contract

Iron law: discover before assuming — never hardcode tool availability; scan the environment first. Invoked once per command run, before any quality or verification phase: /flow:start Phase 1, /flow:pr pre-flight, /flow:address before verification, /flow:resolve step 1c, /flow:setup Phase 1, /flow:design and /flow:brainstorm EXPLORE, and the test-runner agent. Returns the six-section markdown report below (Agents Available, Skills Available, Quality Commands, Tech Stack, Verification Capabilities, LSP Capabilities). The calling command stores the output for later phases and does not re-invoke it within the same run. Permitted skips: the markdown-only early exit (no tech-stack file and no quality commands in CLAUDE.md) and LSP probing when lsp.enabled is false.

Detection Order

Steps 1-5 are independent — run them simultaneously with parallel tool calls.

1. Agents: Glob ".claude/agents/*.md" (project) and "plugins/*/agents/*.md" (plugin); parse source, name (filename without .md), description (frontmatter).

2. Skills: Glob ".claude/skills/*/SKILL.md" and "plugins/*/skills/*/SKILL.md"; parse source, name (parent directory), description.

3. Commands: Glob ".claude/commands/*.md" and "plugins/*/commands/*.md"; parse source, name, description.

4. Quality commands from CLAUDE.md — .claude/CLAUDE.md first, then CLAUDE.md:

   grep -E "(npm|pnpm|yarn|bun|ruff|pytest|go |cargo |make )" "$CLAUDE_MD"
   

5. Tech stack by indicator file: package.json → node, tsconfig.json → typescript, pyproject.toml → python, Gemfile → ruby, go.mod → go, Cargo.toml → rust, Makefile → makefile.

Early exit — markdown-only: no tech-stack file AND no quality commands in CLAUDE.md → report Quality Commands "No code-related quality commands applicable", Tech Stack "Markdown-only project", skip Step 6.

6. Verification capabilities: ls verify.sh scripts/verify* playwright.config.* cypress.config.* 2>/dev/null.

   Also probe the four dependency-license tools (license-checker, pip-licenses, cargo-license, go-licenses) that agents/security-reviewer.md Step 4 uses, and jscpd, the clone detector bin/flow-clone-scan.sh drives. Commands and output lines: references/runtime-verification-probes.md § Tool probes.

   An absent tool is reported not-installed with its install command, never passed over: a reviewer who never learns a tool was missing reads a thinner judgment as a complete one. jscpd has no fallback, so absent it means the verbatim duplication layer reports STATE=unavailable. Report it; never install it.

7. LSP capabilities: pre-check lsp.enabled (default true; if false, report every feature Disabled). Probe documentSymbol, hover, goToDefinition, findReferences, goToImplementation against one representative source file for the detected stack, each bounded by lsp.timeout (default 5000 ms); infer diagnostics Available when documentSymbol succeeds. Procedure and result rules: references/lsp-capability-probes.md.

Output Format

### Agents Available
| Agent | Source | Description |
|-------|--------|-------------|

### Skills Available
| Skill | Source | Description |
|-------|--------|-------------|

### Quality Commands
| Command | Purpose | Source |
|---------|---------|--------|

### Tech Stack
- {language} ({indicator file})

### Verification Capabilities
| Capability | Command | Source |
|-----------|---------|--------|

### LSP Capabilities
| Feature | Status | Use Case |
|---------|--------|----------|
| documentSymbol | {Available/Unavailable} | Symbol navigation in files |
| hover | {Available/Unavailable} | Type info and documentation during CODE phase |
| goToDefinition | {Available/Unavailable} | Trace code paths during EXPLORE phase |
| findReferences | {Available/Unavailable} | Impact analysis during EXPLORE, caller verification during REVIEW |
| goToImplementation | {Available/Unavailable} | Interface resolution |
| diagnostics | {Available/Unavailable} | Quality signal during VERIFY phase (errors→P1, warnings→P2) |

Graceful Degradation

No agents → built-in review checklist. No CLAUDE.md commands → detect from tech stack. No tech stack → ask the user for commands. No LSP server → CLI-only analysis (grep/glob for references, CLI tools for diagnostics); not an error.

Caching

This skill runs in a forked context. The calling command must store the output for use in later phases — do not re-invoke within the same command execution.