claude-bridge

Makes tallow discover and use resources from .claude/ directories, providing compatibility with Claude Code’s file layout conventions. Bridges skills, agents, and commands.

Note

Need the practical compatibility matrix (including hooks and frontmatter caveats)? See Using tallow in existing Claude Code projects.

What gets bridged

The extension enables cross-compatibility with Claude Code by scanning both .tallow/ and .claude/ directories:

Resource	Bridged by	Paths
Skills	claude-bridge	~/.claude/skills/, .claude/skills/
Agents	subagent-tool, agent-commands-tool, context-fork	~/.claude/agents/, .claude/agents/
Commands	command-prompt	~/.claude/commands/, .claude/commands/
Hooks	hooks extension (not claude-bridge)	~/.claude/settings.json, .claude/settings.json

Skill bridging

This extension hooks resources_discover to inject .claude/skills/ paths into the pi framework’s skill loader. Discovered skills:

• Appear in the system prompt’s skill descriptions
• Are callable via /skill:name commands
• Work identically to .tallow/skills/ skills

Paths checked:

1. ~/.claude/skills/ — user-level
2. .claude/skills/ — project-level (relative to cwd)

The pi framework merges these with .tallow/skills/ paths from built-in discovery.

Agent and command bridging

Agents and commands use inline bridging — each extension that loads these resources scans both .tallow/ and .claude/ directories directly:

• Agents: subagent-tool, agent-commands-tool, context-fork
• Commands: command-prompt

This is different from skills because agents and commands are extension-specific, not framework-level resources.

Priority on name collision

When a resource exists in both .tallow/ and .claude/, .tallow/ wins:

.tallow/skills/researcher/    ← loaded (priority)
.claude/skills/researcher/    ← skipped (collision)

This applies at each scope level:

• User level: ~/.tallow/skills/ > ~/.claude/skills/
• Project level: .tallow/skills/ > .claude/skills/

The priority is implemented via load order — .claude/ paths are loaded first, then .tallow/ paths overwrite any duplicate names in the map.

Why separate extension for skills?

Skills use the resources_discover hook to integrate with the pi framework’s built-in skill loader. This needs to run before skills are processed during session initialization.

Agents and commands are extension-owned resources, so each extension handles .claude/ scanning directly in its own discovery logic.

Directory structure example

~/
├── .claude/
│   ├── skills/
│   │   └── web-search/
│   ├── agents/
│   │   └── researcher.md
│   └── commands/
│       └── analyze.md
└── .tallow/
    ├── skills/
    │   └── code-review/
    ├── agents/
    │   └── tester.md
    └── commands/
        └── refactor.md

project/
├── .claude/
│   ├── skills/
│   │   └── project-utils/
│   ├── agents/
│   │   └── deployer.md
│   └── commands/
│       └── build.md
└── .tallow/
    ├── skills/
    │   └── test-runner/
    ├── agents/
    │   └── reviewer.md
    └── commands/
        └── deploy.md

All resources from both .claude/ and .tallow/ directories are available. If both have a resource with the same name, .tallow/ version is used.

Relationships with other extensions

skill-commands

claude-bridge enhances skill-commands by providing .claude/skills/ paths via the resources_discover hook. skill-commands processes these paths the same way it handles .tallow/skills/ paths.

subagent-tool

subagent-tool handles .claude/agents/ bridging directly in its agent discovery logic, not via this extension.

agent-commands-tool

agent-commands-tool handles .claude/agents/ bridging directly, same as subagent-tool.

context-fork

context-fork scans .claude/agents/ alongside .tallow/agents/ when resolving agent configurations for fork commands.

command-prompt

command-prompt scans .claude/commands/ alongside .tallow/commands/ when building the command index.