scoped-agent-control 0.2.0

Role-based, scoped AI query/edit operations for any codebase.

scoped-agent-control

scoped-agent-control is a framework for enforcing role-based, scoped AI interactions with a codebase.

It allows teams to define exactly who can read or modify which parts of a repository and under what constraints. Simple inline annotations and a repo-level config enable AI agents (e.g. Claude Code, Codex, etc.) to operate within approved, role-specific boundaries.

In practice, this enables non-technical users (e.g. PMs, GTM, etc.) to safely make targeted changes without risking unintended or disallowed side effects across the system.

Install

pipx install git+https://github.com/Shriram-Vasudevan/scoped-agent-control
# or
uv tool install git+https://github.com/Shriram-Vasudevan/scoped-agent-control

Both put scoped-control on PATH globally. Requires Python 3.12+.

For local development:

git clone https://github.com/Shriram-Vasudevan/scoped-agent-control
cd scoped-agent-control
uv sync --dev
uv run scoped-control

Quick start

cd path/to/any/repo
scoped-control

The TUI opens. On first run, a wizard asks four questions — role name, description, what to read, what to edit. Press Enter to accept defaults. Setup writes .scoped-control/config.yaml and drops into the main console.

Commands

Type slash commands at the TUI prompt:

Command	Purpose
/setup (or /init)	Guided role wizard
/role list | add | edit | remove	Manage roles
/surface list | show <id>	Inspect indexed surfaces
/scan	Rebuild the index
/query <role> <request>	Scoped read
/edit <role> <request>	Scoped edit with enforcement
/annotate --role <name>	Add file-scope annotations
/cleanup --dry-run | --force	Remove all artifacts
/install claude-code | github | slack	Install integrations
/help, /clear, /quit	TUI controls

The same commands work as CLI subcommands without the slash:

scoped-control role list
scoped-control query maintainer "explain the auth flow"
scoped-control edit recruiter "update the careers intro" --executor codex

CLI-only commands (long-running or guided):

Command	Purpose
triage <request>	Classify a request without running it
install slack-bot	Guided Slack bot setup with auto-tunneling
serve-slack	Run the Slack bot server
remote-edit --event-file <path>	Process a GitHub remote-edit payload
remote-triage --event-file <path>	Process a GitHub remote-triage payload

End-to-end example

$ cd ~/my-repo
$ scoped-control
# Wizard:
#   Role name → recruiter
#   Description → Recruiting copy editor
#   Read → careers folder and hiring docs
#   Write → only careers/**/*.md
# Output: Role `recruiter` created.

# In the TUI prompt:
> /query recruiter what does the careers intro say
> /edit recruiter update the careers intro to mention remote work

# Or from a shell:
$ scoped-control triage "refactor src/auth.py"
Triage decision: mode=blocked role=None
Reason: No configured role has edit access to the requested target(s).

Slack bot

scoped-control install slack-bot

A guided five-step flow:

1. Starts a public tunnel (cloudflared preferred, ngrok fallback).
2. Prints a Slack app manifest with the URL pre-filled.
3. Prompts for the Signing Secret.
4. Waits for "Install to Workspace" in the Slack admin UI.
5. Smoke-tests the round trip and prints a team announcement template.

Total time: ~3 minutes. Requires cloudflared or ngrok on PATH. The tunnel URL is valid only while the process runs; production deployments need a stable domain.

GitHub workflows

scoped-control install github

Writes two workflow_dispatch workflows:

• scoped-control.yml — explicit role + request.
• scoped-control-triage.yml — triage picks the role and mode from a request alone.

Add ANTHROPIC_API_KEY and/or OPENAI_API_KEY as Actions secrets before running.

Claude Code slash commands

scoped-control install claude-code

Writes .claude/commands/sc-{setup,triage,query,edit,surfaces}.md. The slash commands work inside any Claude Code session in the repo.

Annotations

Setup writes config only; no files are modified. Files matching a role's globs are accessible without per-file annotations — the resolver synthesizes whole-file surfaces on demand.

Annotations are inline overrides for two cases:

• Carve-outs — restrict one file to a narrower role than its directory's role.
• Span rules — invariants: span_scope limits edits to a specific function or class.

# surface: core.compute_total
# roles: maintainer
# modes: query, edit
# invariants: span_scope
def compute_total(items):
    ...

Fields: surface, roles, modes, invariants, depends_on. # and // comment styles supported.

Apply with /annotate --role <name> or scoped-control setup --annotate-files.

Enforcement

Edits run in a sandbox workspace and pass through:

• Scope — target file matches edit_paths.
• Span — modifications stay within the target surface's line range.
• Diff limit — limits.max_changed_files, limits.max_diff_lines.
• Validators — commands listed under validators: with mode edit.

Failures roll back the sandbox and surface the reason.

Configuration

.scoped-control/config.yaml:

roles:
  - name: recruiter
    description: Recruiting copy editor
    query_paths: ["careers/**"]
    edit_paths: ["careers/**/*.md"]

limits:
  max_changed_files: 5
  max_diff_lines: 400

validators:
  - name: lint
    command: "ruff check"
    modes: [edit]

Testing

uv run pytest