agent-boundary-kit
Advanced tools
Open-source checks for AI coding-agent boundary failures and native agent integrations.
Agent Boundary Kit is a research-first open-source tooling repo for preventing recurring AI coding-agent boundary failures:
The agent solves the wrong problem while producing output that looks plausible.
That boundary failure shows up as copied internal brief text, negative constraints leaking into UI copy, fallback code added before diagnosis, tests changed only to pass, oversized plans accepted without phase gates, and completion claims without evidence.
The current focus is not selling a plugin. The focus is proving the failure model: taxonomy -> reproducible fixture -> pass/fail rubric -> red/green evidence -> scanner or evaluator. The Codex and Claude plugin candidates are distribution surfaces for proven checks, not the center of the project.
This repo turns those failures into neutral fixtures, pass/fail rubrics, scanner checks, agent instruction templates, and native integration candidates for coding agents.
User direction: "Do not make this sound corporate or salesy."
Bad agent output:
"This is not corporate, not salesy, and not enterprise-sounding."
ABK result:
fail - negative constraint leaked into final copy.
The same boundary shows up in code work:
User direction: "The fallback is wrong. Find the root cause."
Bad agent behavior: adds another fallback.
ABK result: fail - fallback over root cause.
ABK is adjacent to Devflow Native, but it should not own the same layer.
Use Devflow to remember and resume work. Use ABK to stop a plausible-looking but wrong agent move before it becomes code, tests, docs, or a completion claim.
This is not a prompt collection, a dashboard, or a general agent-management app.
It is a kit for:
The research program is defined in docs/research-program.md. New work should start from a failure seed or evidence gap, not from plugin UX polish. Concrete case studies are recorded in docs/case-study-research-mode-no-write.md and docs/case-study-test-passing-not-merge-worthy.md.
Private examples can be used as research seeds only after they are neutralized: remove personal details, preserve the failure shape, and define observable pass/fail criteria.
User input has roles:
A passing agent classifies the role before writing public text, editing code, changing tests, or claiming completion.
The current taxonomy covers:
Runnable fixtures live under benchmarks/fixtures. Each fixture is a small broken repo with a prompt, trap, expected result, verifier, and source notes.
Repository checks:
npm run bench:check
npm run bench:check:red
See docs/benchmarks.md for the benchmark system, runner commands, scanner coverage, and publication rules.
The intended path is agent-native review: open Codex or Claude Code in the target repo and ask it to install Agent Boundary Kit safely.
Install Agent Boundary Kit for this repository.
Inspect the repo first. Preserve existing AGENTS.md, CLAUDE.md, README, tests,
hooks, local settings, and project rules. Use npx agent-boundary-kit@latest if
the package is not already installed.
Run a dry-run first. Show me the runner input you plan to use before running a
scanner. Do not pass private transcripts, hidden chat history, broad workspace
dumps, cookies, tokens, or unreviewed user examples.
If Codex or Claude Code integration is useful, review the candidate skill,
plugin, MCP, or hook files first. Do not edit my persistent Codex or Claude Code
settings unless I explicitly approve the exact configuration change.
Run the relevant ABK checks and tell me exactly what files changed, what scanner
evidence was produced, and what I still need to apply manually.
For manual first use without agent setup:
npx agent-boundary-kit@latest harness inspect
npx agent-boundary-kit@latest harness plan
npx agent-boundary-kit@latest dry-run --input runner-input.json
npx agent-boundary-kit@latest scan --input runner-input.json --scanner legacy-surface-retention-scan
The simplest path is direct local execution:
npx agent-boundary-kit harness inspect
npx agent-boundary-kit dry-run --input runner-input.json
npx agent-boundary-kit scan --input runner-input.json --scanner legacy-surface-retention-scan
For repeated use:
npm install -g agent-boundary-kit
agent-boundary-kit harness inspect
abk-runner dry-run --input runner-input.json
abk-runner scan --input runner-input.json --scanner legacy-surface-retention-scan
Runner input must be explicit. Do not pass private transcripts, hidden chat history, broad workspace dumps, cookies, tokens, or unreviewed user examples. The runner is meant to check declared files and metadata, then return evidence.
Prefer the Codex plugin when you want ABK available across repositories without copying per-repo skills or MCP config. The package includes a repo marketplace at .agents/plugins/marketplace.json and a Codex plugin at plugins/codex-agent-boundary-kit.
agent-boundary-kit harness inspect
agent-boundary-kit harness install --confirm
harness install --confirm registers the GitHub marketplace with the official Codex CLI command:
codex plugin marketplace add Sungblab/agent-boundary-kit
Then restart Codex, open Plugins in the Codex app or /plugins in Codex CLI, install Agent Boundary Kit, and start a new thread. Plugin install and hook trust remain user-reviewed Codex steps.
Use the package through the plugin, CLI, or shared MCP server:
npx agent-boundary-kit ... or abk-runner ... from the repository being checked.abk-mcp-server as a stdio MCP server when you want list_scanners, validate_runner_input, dry_run, and scan exposed as tools.Codex may review the candidate files, explain the exact config change, and run repository evidence gates. The user owns any persistent Codex configuration change.
Use the package through the CLI, the shared MCP server, or the reviewable Claude Code plugin candidate:
npx agent-boundary-kit ... or abk-runner ... from the repository being checked.abk-mcp-server when you want ABK scanner tools available in Claude Code.Claude Code may review the candidate, generate a review packet, and explain the expected user-owned configuration action. Hook and plugin enablement remains a user-approved configuration step.
The package exposes these binaries:
agent-boundary-kit: alias for abk-runner.abk-runner: maps explicit runner input to dry-run and read-only scanner execution; also exposes harness inspect, harness plan, harness install, and harness health for plugin readiness.abk-mcp-server: exposes list_scanners, validate_runner_input, dry_run, and scan for Codex, Claude Code, and MCP-compatible clients.abk-claude-hook: maps explicit Claude hook event envelopes to runner input.abk-claude-hook-wrapper: wraps native Claude hook payloads with explicit ABK carrier metadata.The MCP contract is docs/mcp-server-contract.md. It keeps scanner output as evidence, not final copy.
The repository includes reviewable native integration candidates:
These candidates package the boundary skill and shared abk-mcp-server configuration. The Codex candidate is exposed through the repo marketplace so users can install it once from Codex instead of copying files into each repository. The candidates do not apply user hook settings automatically.
The candidates are review targets, not automatic setup instructions. Keep user-owned Codex and Claude Code configuration separate from this repository until the user explicitly applies a reviewed configuration change.
Before each npm release, run:
Run:
npm run bench:check
npm run bench:check:red
npm run pack:dry-run
Do not run npm publish for a new version until package contents, docs, and integration candidates have been reviewed from the dry-run output.
The repo is moving from research seed to open-source productization. It already contains runnable fixtures, scanner-backed checks, public case candidates, boundary templates, manual packaging contracts, and local runner commands.
The product target is not a dashboard or SaaS workflow. It is native agent integration: Codex skill/plugin/MCP/hook surfaces, Claude Code plugin/skill/MCP/hook surfaces, and a shared read-only MCP server contract backed by the existing benchmark evidence.
The detailed benchmark and hook contracts are kept out of the main overview:
The agent should not ask only, "What words did the user say?"
It should ask:
What role did this input play, and what output would satisfy that role without leaking it?
MIT. See LICENSE.
FAQs
Open-source checks for AI coding-agent boundary failures and native agent integrations.
We found that agent-boundary-kit demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
NIST disclosed an unreleased AI tool called V-etalon and opened a broad inquiry into NVD modernization after years of automation plans produced no public enrichment system.
Security News
In his AI Council 2026 talk, Feross Aboukhadijeh covers recent package compromises, vulnerability discovery, and a more automated security model.
Security News
A new federal program will let vetted U.S. cybersecurity firms help investigate and disrupt foreign cybercrime groups under government direction.