code-agnostic 0.3.4__tar.gz → 0.3.7__tar.gz

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-agnostic
-Version: 0.3.4
+Version: 0.3.7
 Summary: Centralized hub for LLM coding config: MCP, skills, rules, and agents.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -36,20 +36,29 @@ AI coding tools each want config in a different place and format. When you use m
 ├── config/
 │   └── mcp.base.json             MCP servers (editor-agnostic)
 ├── rules/
-│   └── python-style.md           Rules with YAML frontmatter
+│   └── python-style/
+│       ├── meta.yaml             Rule metadata
+│       └── prompt.md             Rule instructions
 ├── skills/
-│   └── code-reviewer/SKILL.md    Skills with YAML frontmatter
+│   └── code-reviewer/
+│       ├── meta.yaml             Skill metadata
+│       └── prompt.md             Skill instructions
 └── agents/
-    └── architect.md              Agents with YAML frontmatter
+    └── architect/
+        ├── meta.yaml             Agent metadata
+        └── prompt.md             Agent instructions
 
         ↓ plan / apply ↓
 
 ~/.config/opencode/               Compiled & synced for OpenCode
 ~/.cursor/                        Compiled & synced for Cursor
 ~/.codex/                         Compiled & synced for Codex
+~/.claude.json and ~/.claude/      Compiled & synced for Claude Code
 ```
 
-Each resource is cross-compiled to the target editor's native format. Rules become `.mdc` files for Cursor, `AGENTS.md` sections for OpenCode/Codex, etc.
+Each resource is cross-compiled to the target editor's native format. Rules become `.mdc` files for Cursor, `AGENTS.md` sections for OpenCode/Codex, and `CLAUDE.local.md` memory for Claude Code.
+
+Legacy single-file rules, `skills/<name>/SKILL.md`, and markdown agents are still supported for migration, but bundle directories are the preferred source format for new config.
 
 Today the implementation is still mixed: some assets are compiled and some are symlinked. The active migration plan is to move to generated outputs everywhere with a strict compiler contract instead of implicit per-app behavior.
 
@@ -86,33 +95,35 @@ code-agnostic import apply -a codex
 # Enable target editors
 code-agnostic apps enable -a cursor
 code-agnostic apps enable -a opencode
+code-agnostic apps enable -a claude
 
 # Preview and apply
+code-agnostic validate
 code-agnostic plan
 code-agnostic apply
 ```
 
 ## Editor compatibility
 
-| Feature | OpenCode | Cursor | Codex |
-|---------|:--------:|:------:|:-----:|
-| MCP sync | yes | yes | yes |
-| Rules sync (cross-compiled) | yes | yes | yes |
-| Skills sync | yes | yes | yes |
-| Agents sync | yes | yes | yes |
-| Workspace root `AGENTS.md` link | yes | yes | yes |
-| Native repo config include for workspace `AGENTS.md` | yes | -- | -- |
-| Repo/subdir gets shared workspace `AGENTS.md` today | yes | -- | yes |
-| Root-level `AGENTS.md` discovery only | -- | yes | yes |
-| Workspace propagation | yes | -- | yes |
-| Import from | yes | yes | yes |
-| Interactive import (TUI) | yes | yes | yes |
+| Feature | OpenCode | Cursor | Codex | Claude Code |
+|---------|:--------:|:------:|:-----:|:-----------:|
+| MCP sync | yes | yes | yes | yes |
+| Rules sync (cross-compiled) | yes | yes | yes | yes |
+| Skills sync | yes | yes | yes | yes |
+| Agents sync | yes | yes | yes | yes |
+| Workspace root `AGENTS.md` link | yes | yes | yes | yes |
+| Native repo config include for workspace `AGENTS.md` | yes | -- | -- | -- |
+| Repo/subdir gets shared workspace instructions today | yes | -- | yes | yes |
+| Nested `AGENTS.md` discovery | -- | yes | yes | -- |
+| Workspace propagation | yes | -- | yes | yes |
+| Import from | yes | yes | yes | yes |
+| Interactive import (TUI) | yes | yes | yes | yes |
 
 Cursor workspace propagation is intentionally disabled to avoid duplicate MCP initialization in multi-root workspaces: https://forum.cursor.com/t/mcp-multi-root-workspace-causes-duplicate-mcp-server-initialization-4x-createclient-actions/144003
 
-OpenCode workspace configs include the shared workspace `AGENTS.md` natively via `instructions`, so repos under the workspace get both repo-local and shared workspace instructions. Codex repos receive workspace instructions through a generated `AGENTS.override.md`, which is added to each repo's `.git/info/exclude`.
+OpenCode workspace configs include the shared workspace `AGENTS.md` natively via `instructions`, so repos under the workspace get both repo-local and shared workspace instructions. Codex repos receive workspace instructions through a generated `AGENTS.override.md`, which is added to each repo's `.git/info/exclude`. Claude Code receives workspace instructions through generated `CLAUDE.local.md` files, never by editing committed `CLAUDE.md`.
 
-Cursor documents `AGENTS.md` as a root-level project file. Codex documents `AGENTS.md` discovery, but not a native config include for an extra workspace file. In practice that means subdirectories and repos opened below the workspace root will not reliably get the shared workspace `AGENTS.md` today for Cursor.
+Cursor documents `AGENTS.md` support in project roots and subdirectories. `code-agnostic` still disables Cursor workspace propagation, so it does not copy or link the shared workspace `AGENTS.md` into child repos; Cursor will load repo-local or nested `AGENTS.md` files that already exist in the opened project. Codex documents nested `AGENTS.md` discovery, but not a native config include for an extra workspace file.
 
 ## Features
 
@@ -121,12 +132,20 @@ Cursor documents `AGENTS.md` as a root-level project file. Codex documents `AGEN
 Plan-then-apply workflow. Preview every change before it touches disk.
 
 ```bash
+code-agnostic validate              # check canonical source files
 code-agnostic plan -a cursor        # dry-run for one editor
 code-agnostic plan                   # dry-run for all
 code-agnostic apply                  # apply changes
 code-agnostic status                 # check drift
 ```
 
+If managed outputs need repair after an apply, restore the active synced revision:
+
+```bash
+code-agnostic restore
+code-agnostic restore -w myproject
+```
+
 ### MCP management
 
 Add, remove, and list MCP servers without editing JSON by hand.
@@ -162,16 +181,37 @@ code-agnostic rules remove --name python-style
 
 ### Skills and agents
 
-Canonical YAML frontmatter format, cross-compiled per editor.
+Canonical YAML frontmatter format, cross-compiled per editor. Install or edit skills in the `code-agnostic` source of truth, then run `plan` / `apply`; do not hand-copy generated skills into `.codex`, `.cursor`, or OpenCode directories.
 
 ```bash
 code-agnostic skills list
 code-agnostic agents list
 ```
 
+Manual skill install today:
+
+```bash
+mkdir -p ~/.config/code-agnostic/skills
+cp -R ./my-skill ~/.config/code-agnostic/skills/my-skill
+code-agnostic plan
+code-agnostic apply
+```
+
+Global skills live under `~/.config/code-agnostic/skills`. Workspace-local skills live under `~/.config/code-agnostic/workspaces/<name>/skills` and can be inspected with `code-agnostic skills list -w <name>`. Codex generated skill outputs are written to `~/.agents/skills`, while Codex agents and config remain under `~/.codex`. Claude Code generated skills and agents are written under `~/.claude/skills` and `~/.claude/agents`, with workspace copies under repo-local `.claude/skills` and `.claude/agents`.
+
+Project-local skills are not first-class source inputs in `code-agnostic` yet. If a target app discovers repo-local skill folders such as `.agents/skills`, `.opencode/skills`, or user-created `.claude/skills`, treat those as unmanaged app inputs. Workspace sync writes only the exact generated paths recorded in `.sync-state.json`.
+
+Planned convenience command:
+
+```bash
+code-agnostic skills install ./my-skill --apply
+```
+
+That command should copy the skill into the source of truth and then run the normal compiler/apply flow.
+
 ### Workspaces
 
-Register workspace directories. Workspace rules are compiled into a canonical `AGENTS.md` at the workspace root. Repos keep their own repo-specific `AGENTS.md`; Codex receives the workspace rules through generated, git-excluded `AGENTS.override.md` files, while OpenCode workspace configs reference the shared workspace file through `instructions`. Repo-local app config, skills, and agents are propagated for OpenCode and Codex.
+Register workspace directories. Workspace rules are compiled into a canonical `AGENTS.md` at the workspace root. Repos keep their own repo-specific `AGENTS.md`; Codex receives the workspace rules through generated, git-excluded `AGENTS.override.md` files, while OpenCode workspace configs reference the shared workspace file through `instructions`. Claude receives generated `CLAUDE.local.md` files and project MCP entries in `~/.claude.json["projects"][absolute_repo_path]["mcpServers"]`. Repo-local app config, skills, and agents are propagated for OpenCode, Codex, and Claude.
 
 `.cursor` workspace propagation is intentionally disabled to avoid duplicate MCP initialization when opening multi-root workspaces in Cursor (related issue: https://forum.cursor.com/t/mcp-multi-root-workspace-causes-duplicate-mcp-server-initialization-4x-createclient-actions/144003).
 
@@ -198,6 +238,7 @@ Migrate existing config from any supported editor into the hub.
 ```bash
 code-agnostic import plan -a codex
 code-agnostic import apply -a codex
+code-agnostic import plan -a claude
 code-agnostic import apply -a cursor --include mcp --on-conflict overwrite
 code-agnostic import plan -a codex -i    # interactive TUI picker
 ```
@@ -210,13 +251,12 @@ All commands use named flags (`-a`, `-w`, `-v`). Singular aliases work too: `app
 
 The compiler migration is documented in:
 
-- [ROADMAP.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/ROADMAP.md)
-- [docs/compiler/overview.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/overview.md)
-- [docs/compiler/skills.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/skills.md)
-- [docs/compiler/agents.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/agents.md)
-- [docs/compiler/rules.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/rules.md)
-- [docs/compiler/mcp.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/mcp.md)
-- [docs/compiler/lossiness.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/lossiness.md)
+- [docs/compiler/overview.md](docs/compiler/overview.md)
+- [docs/compiler/skills.md](docs/compiler/skills.md)
+- [docs/compiler/agents.md](docs/compiler/agents.md)
+- [docs/compiler/rules.md](docs/compiler/rules.md)
+- [docs/compiler/mcp.md](docs/compiler/mcp.md)
+- [docs/compiler/lossiness.md](docs/compiler/lossiness.md)
 
 ## Roadmap
 
@@ -231,7 +271,8 @@ The compiler migration is documented in:
 - [x] Cross-compilation for skills and agents
 - [x] Per-workspace git-exclude customization
 - [x] Interactive TUI for import selection
-- [ ] Claude Code support
+- [x] Claude Code support
+- [ ] Project-scoped skill installs and sync
 - [ ] `rules add` / `skills add` / `agents add` commands (open `$EDITOR` with template)
 - [ ] Planner integration for cross-compiled skills and agents
 - [ ] Shell auto-complete
@@ -243,3 +284,38 @@ The compiler migration is documented in:
 uv sync --dev
 uv run pytest
 ```
+
+Before pushing release-prep work, run the supported Python matrix locally. Tox
+delegates each environment to `uv run --python`, so `uv` can provide the
+requested interpreter when it is not already installed:
+
+```bash
+uvx tox run -p auto
+uvx tox run -e uv310 -- tests/test_version.py -q
+```
+
+For a hermetic Linux matrix that does not depend on locally installed Python
+versions, run the Docker matrix:
+
+```bash
+./scripts/run-docker-matrix.sh
+./scripts/run-docker-matrix.sh tests/test_version.py -q
+```
+
+Limit the Docker matrix while iterating:
+
+```bash
+PYTHON_VERSIONS="3.10 3.14" ./scripts/run-docker-matrix.sh tests/test_version.py -q
+```
+
+The release gate still requires GitHub Actions to pass because the published
+workflow is the source of truth for OS coverage across Ubuntu, macOS, and
+Windows on every supported Python version.
+
+Real app-ingestion E2E is gated because it requires installed target CLIs and
+uses each tool's own introspection surface:
+
+```bash
+CODE_AGNOSTIC_REAL_APP_E2E=1 uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q
+CODE_AGNOSTIC_REAL_APP_E2E=1 CODE_AGNOSTIC_REAL_APP_TARGETS=codex,opencode,claude uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q
+```

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/README.md

@@ -13,20 +13,29 @@ AI coding tools each want config in a different place and format. When you use m
 ├── config/
 │   └── mcp.base.json             MCP servers (editor-agnostic)
 ├── rules/
-│   └── python-style.md           Rules with YAML frontmatter
+│   └── python-style/
+│       ├── meta.yaml             Rule metadata
+│       └── prompt.md             Rule instructions
 ├── skills/
-│   └── code-reviewer/SKILL.md    Skills with YAML frontmatter
+│   └── code-reviewer/
+│       ├── meta.yaml             Skill metadata
+│       └── prompt.md             Skill instructions
 └── agents/
-    └── architect.md              Agents with YAML frontmatter
+    └── architect/
+        ├── meta.yaml             Agent metadata
+        └── prompt.md             Agent instructions
 
         ↓ plan / apply ↓
 
 ~/.config/opencode/               Compiled & synced for OpenCode
 ~/.cursor/                        Compiled & synced for Cursor
 ~/.codex/                         Compiled & synced for Codex
+~/.claude.json and ~/.claude/      Compiled & synced for Claude Code
 ```
 
-Each resource is cross-compiled to the target editor's native format. Rules become `.mdc` files for Cursor, `AGENTS.md` sections for OpenCode/Codex, etc.
+Each resource is cross-compiled to the target editor's native format. Rules become `.mdc` files for Cursor, `AGENTS.md` sections for OpenCode/Codex, and `CLAUDE.local.md` memory for Claude Code.
+
+Legacy single-file rules, `skills/<name>/SKILL.md`, and markdown agents are still supported for migration, but bundle directories are the preferred source format for new config.
 
 Today the implementation is still mixed: some assets are compiled and some are symlinked. The active migration plan is to move to generated outputs everywhere with a strict compiler contract instead of implicit per-app behavior.
 
@@ -63,33 +72,35 @@ code-agnostic import apply -a codex
 # Enable target editors
 code-agnostic apps enable -a cursor
 code-agnostic apps enable -a opencode
+code-agnostic apps enable -a claude
 
 # Preview and apply
+code-agnostic validate
 code-agnostic plan
 code-agnostic apply
 ```
 
 ## Editor compatibility
 
-| Feature | OpenCode | Cursor | Codex |
-|---------|:--------:|:------:|:-----:|
-| MCP sync | yes | yes | yes |
-| Rules sync (cross-compiled) | yes | yes | yes |
-| Skills sync | yes | yes | yes |
-| Agents sync | yes | yes | yes |
-| Workspace root `AGENTS.md` link | yes | yes | yes |
-| Native repo config include for workspace `AGENTS.md` | yes | -- | -- |
-| Repo/subdir gets shared workspace `AGENTS.md` today | yes | -- | yes |
-| Root-level `AGENTS.md` discovery only | -- | yes | yes |
-| Workspace propagation | yes | -- | yes |
-| Import from | yes | yes | yes |
-| Interactive import (TUI) | yes | yes | yes |
+| Feature | OpenCode | Cursor | Codex | Claude Code |
+|---------|:--------:|:------:|:-----:|:-----------:|
+| MCP sync | yes | yes | yes | yes |
+| Rules sync (cross-compiled) | yes | yes | yes | yes |
+| Skills sync | yes | yes | yes | yes |
+| Agents sync | yes | yes | yes | yes |
+| Workspace root `AGENTS.md` link | yes | yes | yes | yes |
+| Native repo config include for workspace `AGENTS.md` | yes | -- | -- | -- |
+| Repo/subdir gets shared workspace instructions today | yes | -- | yes | yes |
+| Nested `AGENTS.md` discovery | -- | yes | yes | -- |
+| Workspace propagation | yes | -- | yes | yes |
+| Import from | yes | yes | yes | yes |
+| Interactive import (TUI) | yes | yes | yes | yes |
 
 Cursor workspace propagation is intentionally disabled to avoid duplicate MCP initialization in multi-root workspaces: https://forum.cursor.com/t/mcp-multi-root-workspace-causes-duplicate-mcp-server-initialization-4x-createclient-actions/144003
 
-OpenCode workspace configs include the shared workspace `AGENTS.md` natively via `instructions`, so repos under the workspace get both repo-local and shared workspace instructions. Codex repos receive workspace instructions through a generated `AGENTS.override.md`, which is added to each repo's `.git/info/exclude`.
+OpenCode workspace configs include the shared workspace `AGENTS.md` natively via `instructions`, so repos under the workspace get both repo-local and shared workspace instructions. Codex repos receive workspace instructions through a generated `AGENTS.override.md`, which is added to each repo's `.git/info/exclude`. Claude Code receives workspace instructions through generated `CLAUDE.local.md` files, never by editing committed `CLAUDE.md`.
 
-Cursor documents `AGENTS.md` as a root-level project file. Codex documents `AGENTS.md` discovery, but not a native config include for an extra workspace file. In practice that means subdirectories and repos opened below the workspace root will not reliably get the shared workspace `AGENTS.md` today for Cursor.
+Cursor documents `AGENTS.md` support in project roots and subdirectories. `code-agnostic` still disables Cursor workspace propagation, so it does not copy or link the shared workspace `AGENTS.md` into child repos; Cursor will load repo-local or nested `AGENTS.md` files that already exist in the opened project. Codex documents nested `AGENTS.md` discovery, but not a native config include for an extra workspace file.
 
 ## Features
 
@@ -98,12 +109,20 @@ Cursor documents `AGENTS.md` as a root-level project file. Codex documents `AGEN
 Plan-then-apply workflow. Preview every change before it touches disk.
 
 ```bash
+code-agnostic validate              # check canonical source files
 code-agnostic plan -a cursor        # dry-run for one editor
 code-agnostic plan                   # dry-run for all
 code-agnostic apply                  # apply changes
 code-agnostic status                 # check drift
 ```
 
+If managed outputs need repair after an apply, restore the active synced revision:
+
+```bash
+code-agnostic restore
+code-agnostic restore -w myproject
+```
+
 ### MCP management
 
 Add, remove, and list MCP servers without editing JSON by hand.
@@ -139,16 +158,37 @@ code-agnostic rules remove --name python-style
 
 ### Skills and agents
 
-Canonical YAML frontmatter format, cross-compiled per editor.
+Canonical YAML frontmatter format, cross-compiled per editor. Install or edit skills in the `code-agnostic` source of truth, then run `plan` / `apply`; do not hand-copy generated skills into `.codex`, `.cursor`, or OpenCode directories.
 
 ```bash
 code-agnostic skills list
 code-agnostic agents list
 ```
 
+Manual skill install today:
+
+```bash
+mkdir -p ~/.config/code-agnostic/skills
+cp -R ./my-skill ~/.config/code-agnostic/skills/my-skill
+code-agnostic plan
+code-agnostic apply
+```
+
+Global skills live under `~/.config/code-agnostic/skills`. Workspace-local skills live under `~/.config/code-agnostic/workspaces/<name>/skills` and can be inspected with `code-agnostic skills list -w <name>`. Codex generated skill outputs are written to `~/.agents/skills`, while Codex agents and config remain under `~/.codex`. Claude Code generated skills and agents are written under `~/.claude/skills` and `~/.claude/agents`, with workspace copies under repo-local `.claude/skills` and `.claude/agents`.
+
+Project-local skills are not first-class source inputs in `code-agnostic` yet. If a target app discovers repo-local skill folders such as `.agents/skills`, `.opencode/skills`, or user-created `.claude/skills`, treat those as unmanaged app inputs. Workspace sync writes only the exact generated paths recorded in `.sync-state.json`.
+
+Planned convenience command:
+
+```bash
+code-agnostic skills install ./my-skill --apply
+```
+
+That command should copy the skill into the source of truth and then run the normal compiler/apply flow.
+
 ### Workspaces
 
-Register workspace directories. Workspace rules are compiled into a canonical `AGENTS.md` at the workspace root. Repos keep their own repo-specific `AGENTS.md`; Codex receives the workspace rules through generated, git-excluded `AGENTS.override.md` files, while OpenCode workspace configs reference the shared workspace file through `instructions`. Repo-local app config, skills, and agents are propagated for OpenCode and Codex.
+Register workspace directories. Workspace rules are compiled into a canonical `AGENTS.md` at the workspace root. Repos keep their own repo-specific `AGENTS.md`; Codex receives the workspace rules through generated, git-excluded `AGENTS.override.md` files, while OpenCode workspace configs reference the shared workspace file through `instructions`. Claude receives generated `CLAUDE.local.md` files and project MCP entries in `~/.claude.json["projects"][absolute_repo_path]["mcpServers"]`. Repo-local app config, skills, and agents are propagated for OpenCode, Codex, and Claude.
 
 `.cursor` workspace propagation is intentionally disabled to avoid duplicate MCP initialization when opening multi-root workspaces in Cursor (related issue: https://forum.cursor.com/t/mcp-multi-root-workspace-causes-duplicate-mcp-server-initialization-4x-createclient-actions/144003).
 
@@ -175,6 +215,7 @@ Migrate existing config from any supported editor into the hub.
 ```bash
 code-agnostic import plan -a codex
 code-agnostic import apply -a codex
+code-agnostic import plan -a claude
 code-agnostic import apply -a cursor --include mcp --on-conflict overwrite
 code-agnostic import plan -a codex -i    # interactive TUI picker
 ```
@@ -187,13 +228,12 @@ All commands use named flags (`-a`, `-w`, `-v`). Singular aliases work too: `app
 
 The compiler migration is documented in:
 
-- [ROADMAP.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/ROADMAP.md)
-- [docs/compiler/overview.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/overview.md)
-- [docs/compiler/skills.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/skills.md)
-- [docs/compiler/agents.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/agents.md)
-- [docs/compiler/rules.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/rules.md)
-- [docs/compiler/mcp.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/mcp.md)
-- [docs/compiler/lossiness.md](/Users/alexeyartishevsky/PycharmProjects/llm-sync/docs/compiler/lossiness.md)
+- [docs/compiler/overview.md](docs/compiler/overview.md)
+- [docs/compiler/skills.md](docs/compiler/skills.md)
+- [docs/compiler/agents.md](docs/compiler/agents.md)
+- [docs/compiler/rules.md](docs/compiler/rules.md)
+- [docs/compiler/mcp.md](docs/compiler/mcp.md)
+- [docs/compiler/lossiness.md](docs/compiler/lossiness.md)
 
 ## Roadmap
 
@@ -208,7 +248,8 @@ The compiler migration is documented in:
 - [x] Cross-compilation for skills and agents
 - [x] Per-workspace git-exclude customization
 - [x] Interactive TUI for import selection
-- [ ] Claude Code support
+- [x] Claude Code support
+- [ ] Project-scoped skill installs and sync
 - [ ] `rules add` / `skills add` / `agents add` commands (open `$EDITOR` with template)
 - [ ] Planner integration for cross-compiled skills and agents
 - [ ] Shell auto-complete
@@ -220,3 +261,38 @@ The compiler migration is documented in:
 uv sync --dev
 uv run pytest
 ```
+
+Before pushing release-prep work, run the supported Python matrix locally. Tox
+delegates each environment to `uv run --python`, so `uv` can provide the
+requested interpreter when it is not already installed:
+
+```bash
+uvx tox run -p auto
+uvx tox run -e uv310 -- tests/test_version.py -q
+```
+
+For a hermetic Linux matrix that does not depend on locally installed Python
+versions, run the Docker matrix:
+
+```bash
+./scripts/run-docker-matrix.sh
+./scripts/run-docker-matrix.sh tests/test_version.py -q
+```
+
+Limit the Docker matrix while iterating:
+
+```bash
+PYTHON_VERSIONS="3.10 3.14" ./scripts/run-docker-matrix.sh tests/test_version.py -q
+```
+
+The release gate still requires GitHub Actions to pass because the published
+workflow is the source of truth for OS coverage across Ubuntu, macOS, and
+Windows on every supported Python version.
+
+Real app-ingestion E2E is gated because it requires installed target CLIs and
+uses each tool's own introspection surface:
+
+```bash
+CODE_AGNOSTIC_REAL_APP_E2E=1 uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q
+CODE_AGNOSTIC_REAL_APP_E2E=1 CODE_AGNOSTIC_REAL_APP_TARGETS=codex,opencode,claude uv run pytest tests/e2e/test_real_app_ingestion_e2e.py -q
+```

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/code_agnostic/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.1.0"
+__version__ = "0.3.7"

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/code_agnostic/__main__.py

@@ -48,13 +48,15 @@ cli.add_command(import_group)
 
 def main() -> int:
     try:
-        cli(standalone_mode=False)
+        result = cli(standalone_mode=False)
     except click.exceptions.Exit as exc:
         code = exc.exit_code
         return code if isinstance(code, int) else 1
     except click.ClickException as exc:
         exc.show()
         return 2
+    if isinstance(result, int):
+        return result
     return 0
 
 

code_agnostic-0.3.7/code_agnostic/agents/claude.py

@@ -0,0 +1,67 @@
+"""Claude Code subagent Markdown conversion."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import Any
+
+import yaml  # type: ignore[import-untyped]
+
+from code_agnostic.agents.models import Agent
+
+_SAFE_FILE_STEM_RE = re.compile(r"[^A-Za-z0-9_-]+")
+
+
+def normalize_claude_agent_filename(name: str, fallback: str) -> str:
+    candidate = name.strip() or fallback.strip()
+    normalized = _SAFE_FILE_STEM_RE.sub("-", candidate).strip("-_")
+    return normalized or fallback
+
+
+def serialize_claude_agent(agent: Agent) -> str:
+    fm: dict[str, Any] = {}
+    if agent.metadata.name:
+        fm["name"] = agent.metadata.name
+    else:
+        fm["name"] = agent.name
+
+    description = agent.metadata.description or agent.metadata.name or agent.name
+    if description:
+        fm["description"] = description
+
+    model = agent.metadata.effective_value("claude", "model")
+    if model:
+        fm["model"] = model
+    reasoning_effort = agent.metadata.effective_value("claude", "reasoning_effort")
+    if reasoning_effort:
+        fm["effort"] = reasoning_effort
+
+    for key, value in agent.metadata.app_passthrough(
+        "claude",
+        consumed_keys={
+            "model",
+            "reasoning_effort",
+            "sandbox_mode",
+            "nickname_candidates",
+        },
+    ).items():
+        if key in fm:
+            continue
+        fm[key] = value
+
+    parts = [
+        "---",
+        yaml.dump(fm, default_flow_style=False, sort_keys=False).rstrip(),
+        "---",
+        "",
+        agent.content,
+    ]
+    return "\n".join(parts)
+
+
+def claude_agent_target_path(target_dir: Path, agent: Agent) -> Path:
+    return (
+        target_dir
+        / f"{normalize_claude_agent_filename(agent.metadata.name, agent.name)}.md"
+    )

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/code_agnostic/agents/compilers.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from abc import ABC, abstractmethod
 
+from code_agnostic.agents.claude import serialize_claude_agent
 from code_agnostic.agents.codex import serialize_codex_agent
 from code_agnostic.agents.models import Agent
 from code_agnostic.agents.opencode import serialize_opencode_agent
@@ -35,3 +36,10 @@ class CodexAgentCompiler(IAgentCompiler):
 
     def compile(self, agent: Agent) -> str:
         return serialize_codex_agent(agent)
+
+
+class ClaudeAgentCompiler(IAgentCompiler):
+    """Cross-compile for Claude Code subagents."""
+
+    def compile(self, agent: Agent) -> str:
+        return serialize_claude_agent(agent)

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/code_agnostic/agents/opencode.py

@@ -22,25 +22,25 @@ def serialize_opencode_agent(agent: Agent) -> str:
     if reasoning_effort:
         fm["reasoningEffort"] = reasoning_effort
 
-    tools: dict[str, Any] = {}
-    if agent.metadata.tools.read is not True:
-        tools["read"] = agent.metadata.tools.read
-    if agent.metadata.tools.write is not True:
-        tools["write"] = agent.metadata.tools.write
-    if agent.metadata.tools.mcp:
-        tools["mcp"] = agent.metadata.tools.mcp
-    if tools:
-        fm["tools"] = tools
-
-    for key, value in agent.metadata.app_passthrough(
+    passthrough = agent.metadata.app_passthrough(
         "opencode",
         consumed_keys={
             "model",
             "reasoning_effort",
             "sandbox_mode",
             "nickname_candidates",
+            "permission",
         },
-    ).items():
+    )
+
+    permission = _merge_permission(
+        _compile_permission(agent),
+        agent.metadata.app_overrides.get("opencode", {}).get("permission"),
+    )
+    if permission:
+        fm["permission"] = permission
+
+    for key, value in passthrough.items():
         if key in fm:
             continue
         fm[key] = value
@@ -54,3 +54,27 @@ def serialize_opencode_agent(agent: Agent) -> str:
 
     parts.append(agent.content)
     return "\n".join(parts)
+
+
+def _compile_permission(agent: Agent) -> dict[str, Any]:
+    permission: dict[str, Any] = {}
+    permission["read"] = "allow" if agent.metadata.tools.read else "deny"
+    permission["edit"] = "allow" if agent.metadata.tools.write else "deny"
+    for item in agent.metadata.tools.mcp:
+        server = item.get("server")
+        if not server:
+            continue
+        tool = item.get("tool")
+        key = f"{server}_{tool}" if tool else f"{server}_*"
+        permission[key] = "allow"
+    return permission
+
+
+def _merge_permission(
+    generated: dict[str, Any], override: Any | None
+) -> dict[str, Any] | Any:
+    if override is None:
+        return generated
+    if isinstance(override, dict):
+        return {**generated, **override}
+    return override

{code_agnostic-0.3.4 → code_agnostic-0.3.7}/code_agnostic/agents/parser.py

@@ -19,7 +19,7 @@ from code_agnostic.agents.models import (
 from code_agnostic.spec.loaders import load_agent_bundle
 
 _FRONTMATTER_RE = re.compile(r"^---\s*\n(.*?)\n---\s*\n", re.DOTALL)
-_APP_OVERRIDE_PREFIXES = ("cursor", "codex", "opencode")
+_APP_OVERRIDE_PREFIXES = ("cursor", "codex", "opencode", "claude")
 
 
 def parse_agent(path: Path) -> Agent: