AgentLoopKit

A drop-in engineering loop for software agents.

An open-source Baseframe Labs developer tool.

Your software agent can produce a diff. AgentLoopKit helps you decide whether that work is reviewable, verifiable, and merge-ready.

It gives Codex, Claude Code, Cursor, OpenCode, Gemini CLI, GitHub Copilot CLI, and other engineering agents a local acceptance layer: task contracts, safety rules, verification evidence, review-readiness scores, PR descriptions, and reviewer handoffs.

Unstructured automation produces code. Agentic engineering produces auditable work.

ProjScan finds the risk. AgentLoopKit controls the work. AgentFlight proves the result.

The screenshots and terminal demos in this README are generated from committed sources in docs/assets/readme/ with Playwright and VHS. You can regenerate them locally.

What It Does

AgentLoopKit is a repo-level toolkit for developers using Codex, Claude Code, Cursor, OpenCode, Gemini CLI, GitHub Copilot CLI, and other engineering agents.

Keep your agent, IDE, and CLI. AgentLoopKit gives those tools a repeatable local workflow:

1. Specify the task.
2. Constrain the risk.
3. Implement the smallest useful change.
4. Verify with real commands.
5. Review the diff.
6. Hand off with evidence.
7. Ship only when the work is reviewable.

Run init inside an existing repository and AgentLoopKit creates:

• .agentloop/ for tasks, reports, handoffs, gates, policies, templates, and agent instructions
• AGENTS.md for agent-facing repo guidance
• AGENTLOOP.md for the working loop
• agentloop.config.json for local command and path settings

Generated AGENTS.md includes a specialist roster for product, CLI, template, verification, security, release, docs, compatibility, MCP, and repo-steward work. It helps future agent sessions split work without adding a background service.

The CLI demo shows the normal path: initialize a repo, create a task, run reviewed verification, score review readiness, draft PR copy, inspect review context, and trace why a file changed.

Install

Use it with npx from the repository you want to configure:

cd /path/to/your/repo
npx agentloopkit init --dry-run
npx agentloopkit init

init writes files into the current directory. --dry-run previews the same plan and writes nothing.

Commit the generated harness baseline before your first feature task when you want clean diff evidence. If the harness files are still uncommitted, explain-diff, Guard, and resume packs will include those generated files as local work, which is accurate but noisier than the steady-state loop.

After setup, non-init commands search upward for the nearest agentloop.config.json and use that folder as the AgentLoop root. You can run agentloop status, agentloop verify, agentloop ship, or agentloop prepare-pr from a nested source folder and still write tasks, reports, runs, and handoffs to the initialized repo root.

Do not run init from your home directory unless you intend to configure your home folder. AgentLoopKit refuses that by default; --force allows it and prints a home-directory target warning. If you want local agent guidance but do not want to commit the generated files, use local-only mode:

npx agentloopkit init --local-only

--local-only writes the harness, asks Git for this clone's metadata directory, then adds a marked block to that clone's info/exclude for .agentloop/, AGENTS.md, AGENTLOOP.md, and agentloop.config.json. It does not edit .gitignore, global Git config, shell profiles, or other project files.

The examples below use agentloop after setup. If you are not installing the package locally or exposing its binary on PATH, keep using the one-off form, for example npx --yes agentloopkit@latest doctor --redact-paths.

Check the published package explicitly:

npx --yes agentloopkit@latest version
npx --yes agentloopkit@latest init

For repeatable CI or team setup, replace @latest with a vetted version.

Before initialization, exploratory preflight can be advisory:

npx --yes agentloopkit@latest doctor --advisory --redact-paths

Advisory mode keeps missing-setup diagnostics visible but exits 0 so first-run scripts can continue. Use plain doctor or doctor --strict for gates that should fail on setup errors.

Pinned team usage:

pnpm add -D agentloopkit
pnpm agentloop init

Existing Repos

Repos that already use AgentLoopKit can update the CLI without rewriting local harness files:

cd /path/to/existing/repo
npx --yes agentloopkit@latest version
npx --yes agentloopkit@latest doctor --redact-paths
npx --yes agentloopkit@latest upgrade-harness --details --redact-paths
npx --yes agentloopkit@latest init --dry-run

These commands inspect the published package. This repository may document unreleased guidance before the next publish; check release status before expecting agentloopkit@latest to include new subcommands.

init skips existing generated files and appends to unmarked AGENTS.md files instead of overwriting maintainer edits. Use upgrade-harness to find older guidance that does not tell agents to run agentloop doctor and agentloop start before broad reads, list handles with agentloop context handles when available, or expand source truth with agentloop context show <handle>. Copy only the useful pieces into AGENTS.md, AGENTLOOP.md, or .agentloop/harness/*.

You can use the latest loop before refreshing old guidance:

npx --yes agentloopkit@latest create-task --type bugfix --title "Fix checkout bug"
npx --yes agentloopkit@latest start --for generic --goal implement --redact-paths
npx --yes agentloopkit@latest verify
npx --yes agentloopkit@latest ship
npx --yes agentloopkit@latest prepare-pr
npx --yes agentloopkit@latest maintainer-check

Core Workflow

First useful loop after init:

agentloop doctor
agentloop create-task --type bugfix --title "Fix checkout bug" --include-config-commands
agentloop status --brief
agentloop start --for codex --goal implement --redact-paths
agentloop loop create --goal "Keep checkout work release-ready" --budget-tokens 50000 --max-iterations 5
agentloop loop tick
agentloop loop scorecard
agentloop ready
agentloop verify --task-commands --progress
agentloop ship
agentloop prepare-pr
agentloop task done

That sequence creates a task contract, checks the repo setup, gives the next software agent a compact start briefing, records a local loop contract with token receipts, scores whether another loop pass should continue, checks readiness gates, runs only reviewed verification commands, records review-readiness evidence, drafts PR copy, and closes the active task. It does not post to GitHub, publish packages, call an LLM, execute a coding agent, or run hidden commands.

Make Your Agent Start Here

Before a software agent broad-scans the repo, run:

agentloop doctor --redact-paths
agentloop start --for codex --goal implement --redact-paths
agentloop context handles --redact-paths
agentloop context show <handle-from-start> --redact-paths

doctor checks the Agent Readiness Matrix: Doctor-before-Start guidance, Start guidance, context-handle guidance, broad-read avoidance, MCP guidance, and installed agent instruction files. start gives the current task, proof, risk, context budget, and source handles. context handles lists which handles are available now. context show expands exact local source truth only when the agent needs detail.

npx agentloopkit init
agentloop create-task --title "Fix login redirect bug" --type bugfix \
  --acceptance "Password-reset login redirects to the requested page" \
  --include-config-commands \
  --verification "npm test -- auth"

# Run Codex, Claude Code, Cursor, OpenCode, Gemini CLI, or another software agent.

agentloop verify --task-commands --only-task-commands --progress
agentloop ship --github-comment
agentloop prepare-pr

create-task sets the new contract as the active task. Use agentloop task set <path> only when you need to switch to another contract. Add --include-config-commands when you want the task contract to copy non-empty test, lint, typecheck, and build commands from agentloop.config.json. Task creation records those commands; it does not run them. Use verify --task-commands --only-task-commands when the active contract already includes the full check list. Use verify --post-verification-gates only for reviewed task gates that need the verification report to exist first. After verification and ship, use agentloop prepare-pr for reviewer copy and agentloop task done to close the active task before starting unrelated work.

Use agentloop status or agentloop next when an agent needs the next local action without reading every file:

npx agentloopkit status
npx agentloopkit status --brief
npx agentloopkit next
npx agentloopkit next --redact-paths

When status or next recommends create-task in a dirty repo, the reason calls out existing dirty non-evidence files so agents can confirm the next task scope before implementation. When create-task sees that same pre-existing dirty work, it also adds a bounded Risk Notes bullet to the generated task contract so later review evidence preserves the baseline count and examples.

Use agentloop start --for codex --goal implement --redact-paths when a software agent needs the repo truth before broad reads. Start is the preflight: current task when one exists, decisive state, next safe command, read-first handles, risk summary, verification freshness, context-budget impact, and source handles. Use agentloop context handles to list available handles and normal empty states, then agentloop context show <handle> to expand source truth only when needed. Use agentloop context pack --for codex --goal continue --redact-paths when you need the lower-level receipt and omission details directly. Use agentloop guard when you want a live local check for scope drift, stale verification, proof debt, and context-budget pressure before review. Use agentloop explain-diff when you need to understand whether the current changed files are covered by task scope, recent run evidence, fresh verification, and risk guidance. Use agentloop resume-pack --for codex, --for claude, --for cursor, --for generic, or --for human when you need the older compact continuation surface. These commands are read-only by default and use local evidence only.

agentloop start is the agent starting point. It keeps agents from guessing by turning task state, changed files, verification, risk, and context pressure into one compact preflight. The Context Contract sits underneath it: context pack explains why each item is present, names what was left out, and provides local handles for the source truth. A large changed-file list becomes a small briefing the agent can expand only when it needs detail.

agentloop loop adds a local loop contract on top of normal task contracts. agentloop loop create records the goal, budget, stop conditions, suggested commands, native task path, and a token receipt. agentloop loop tick records one iteration decision from local evidence: continue, stop, ask for human review, or mark the loop ready when gates pass. Once a loop asks for human review, AgentLoopKit rejects another tick until someone reviews the loop report and resolves the blocking evidence. Use agentloop ready when you need the current review-readiness gate summary without creating a loop.

For a bounded autonomous pass, configure one local runner command and let AgentLoopKit record the iteration:

agentloop loop create \
  --goal "Keep checkout work release-ready" \
  --runner-command "node scripts/agentloop-runner.mjs" \
  --runner-timeout-ms 600000 \
  --budget-tokens 50000 \
  --max-iterations 5

agentloop loop run
agentloop loop scorecard
agentloop ready --strict
agentloop loop report

loop scorecard is the pre-flight gate before another autonomous pass. It reads the loop contract, readiness state, context pack, token budget, iteration budget, runner guardrails, and scope evidence, then returns continue, ask-human, stop, or ready. It does not execute the runner or run verification.

loop run executes only the command saved in the loop contract. It uses shell: false, rejects shell syntax, blocks publish and destructive command families, caps runner output, records changed files, and still applies the max-iteration and token-budget stop conditions. AgentLoopKit controls the goal, budget, scope, gates, evidence, and stop decision. Your agent, script, CI job, or AgentFlight session does the implementation work.

JSON and MCP context budget/pack payloads keep that boundary: they include compact evidence fields and the evidence-map:current handle, but they do not send the full changed-file list unless the agent expands that local handle.

Start uses a strict current-work rule. It accepts active or open task contracts. Archived, done, deferred, and AgentFlight placeholder tasks stay as previous evidence. Start and Context omit task:active handles and agentloop ship guidance for old work. In a clean repo with only previous evidence, Start tells the agent that no active task exists.

Your software agent / app
  (Codex, Claude Code, Cursor, OpenCode, Gemini CLI, Copilot CLI,
   LangChain, Agno, Strands, your own scripts...)
        |
        | task contracts - diffs - verification - runs - research notes - logs
        v
 +------------------------------------------------------------------------+
 | AgentLoop Start Preflight + Context Contract (source truth stays here) |
 |------------------------------------------------------------------------|
 | State -> Next Safe Command -> Read First -> Risk -> Impact -> Handles   |
 |                                      |                                  |
 |                                      +-- task:active (current work only)|
 |                                      +-- verification:latest            |
 |                                      +-- run:latest                     |
 |                                      +-- evidence-map:current           |
 |                                      +-- context-budget:current         |
 |                                                                        |
 | Goals: implement - continue - review - debug - handoff - research       |
 | MCP tools - generated agent instructions - transparent local heuristics |
 +------------------------------------------------------------------------+
        |
        | compact preflight + retrieval handles
        v
Agent session / review / research handoff

The context-budget example is generated from local AgentLoopKit output in this repository. Token estimates use ceil(character_count / 4) for planning, not provider tokenizer counts or billing claims. The percentage changes with repo state; the point is that agents can start from selected evidence and expand source handles instead of pasting broad history.

Run the repeatable Start usefulness demo to create a temporary messy repo and see agentloop doctor check agent-readiness guidance, then agentloop start report context avoided, broad files avoided, stale proof, scope drift, source handles, and the next safe command.

The terminal demo is generated from committed VHS sources in this repository. It shows the intended agent workflow: check readiness with Doctor, run Start preflight, see what not to broad-scan, expand the current task handle when one exists, and verify with local evidence.

Verification reports turn local commands into review evidence. They keep command status, failure summaries, task links, changed-file context, and follow-up gates in one artifact reviewers can inspect.

Dogfood Gate

AgentLoopKit uses its own loop in this repo. Contributors can run the same local self-check:

npm run dogfood

That command checks task-folder hygiene, current loop status, public docs hygiene, dependency audit results, harness upgrade state, review gates, artifact inventory, maintainer reviewability, agent review context, AgentFlight health, and ProjScan health. It does not publish packages, create tags, post comments, read tokens, read .env contents, or run verification commands.

Use JSON output when an agent or CI job needs a structured dogfood summary:

npm run dogfood:json
npm run dogfood:strict:json

The JSON summary reports the mode, overall status, step results, exit codes, durations, and safety notes. It redacts the current workspace root as [git-root] before printing structured output.

Run the public-doc check by itself when you edit README, examples, or release-channel docs:

npm run check:public-docs

Use the stricter gate before a release or final handoff:

npm run dogfood:strict

Use the maintenance gate when typecheck, release-proof or npm-status command health, public docs, SchemaStore, policy packs, GitHub metadata, AgentFlight, or ProjScan behavior changes:

npm run maintenance:check

maintenance:check covers unit tests, typecheck, public docs, links, non-strict npm release-proof smoke coverage, read-only npm-status safety tests, SchemaStore output and consistency tests, policy-pack inventory and safety tests, GitHub metadata safety and ship-score neutrality tests, AgentFlight, ProjScan, and the non-strict dogfood self-check. Use npm run dogfood:strict after fresh handoff or ship evidence exists, when review-gate warnings should block the handoff. maintainer-check warnings remain reviewer guidance unless the command exits non-zero. Strict public release proof remains part of approved release gates.

Commands

The package exposes two binaries:

agentloop init
agentloopkit init

Clean up finished task contracts after verification and review-readiness evidence:

agentloop task archive --status done --dry-run
agentloop task archive --status done

See docs/cli-reference.md for command examples, JSON modes, and safety notes.

See docs/loop-contracts.md for local loop contracts, readiness gates, token receipts, and dogfood loops.

See Baseframe Suite Integration v1 for the local JSON handoff between ProjScan, AgentLoopKit, and AgentFlight.

Maintainers can run a faster focused sanity set while iterating:

npm run test:quick
npm run test:integration

Run the full suite before release:

npm run test:release
npm run release-flow

test:quick runs the fast unit-oriented set. test:integration runs slower CLI workflow coverage. test:release is the full Vitest suite. release-flow runs the local release gate without publishing packages, creating tags, or creating GitHub releases.

Agent Setup

Install instructions for the agent you use:

agentloop install-agent codex
agentloop install-agent claude-code
agentloop install-agent cursor
agentloop install-agent opencode
agentloop install-agent gemini-cli
agentloop install-agent github-copilot-cli
agentloop install-agent all
agentloop install-agent all --json --redact-paths

AgentLoopKit writes repo-local Markdown instructions. Re-running install-agent is safe: existing .agentloop/agents/<agent>.md files are skipped so local edits are preserved, while AGENTS.md is updated with any missing AgentLoopKit references. Use --redact-paths before sharing setup output; it hides local roots in human and JSON output without changing where files are written. When exact third-party config conventions are uncertain, it keeps the output generic instead of pretending to own another tool's setup.

Safety Model

AgentLoopKit is intentionally boring:

• no postinstall scripts
• no telemetry
• no cloud backend
• no database
• no API keys
• no hidden network calls
• no reading .env contents
• no destructive actions during install

Env files are reported by path only. Verification commands run only when you explicitly run agentloop verify.

agentloop ship reuses current verification evidence by default. Pass --run-verify when you want it to run verification as part of the readiness flow. Pass --github-comment when CI needs compact Markdown for a pull request comment. AgentLoopKit does not post the comment itself.

agentloop prepare-pr groups changed files by review area, including risk-sensitive paths, source, tests, AgentLoop evidence, docs, CI, config, and other files.

agentloop start briefs software agents with the active task, decisive preflight state, next safe command, read-first source handles, risk summary, verification freshness, context-budget impact, and source handles. agentloop context turns local task, diff, verification, and run evidence into a compact context budget or source-handled pack for software agents. It is read-only and explains what was included, what was omitted, why, and how to expand local source truth. agentloop guard adds drift, proof-debt, baseline, and context-budget checks on top of the evidence map. It is read-only unless you explicitly pass --write-report or --write-baseline. agentloop explain-diff shows the evidence map directly. agentloop resume-pack packages the same local evidence into the older compact continuation brief with context-budget estimates. They do not run tests, call external APIs, read .env contents, post comments, publish packages, create tags, intercept prompts, proxy provider traffic, or change task state.

For narrower evidence history, agentloop verify --write-run and agentloop handoff --write-run can add their own local run records without changing the default command behavior.

agentloop status includes the newest local run ledger entry when .agentloop/runs/ exists, so agents can see the latest review-readiness or verification evidence without opening every report. Run ledger output uses safe display paths: .agentloop/... for AgentLoopKit artifacts, repo-relative paths for repo files, and filenames for older outside absolute paths. Use --redact-paths with doctor, task list, task show, task set, task status, task done, task archive, task clear, status, next, start, review-context, context, guard, explain-diff, resume-pack, check-gates, artifacts, report, badge, runs, show-run, intent, verify, summarize, handoff, ship, prepare-pr, maintainer-check, upgrade-harness, ci-summary, release-notes, schemastore, github import, install-agent, release-check, or release-proof before pasting output into a public issue, PR, or CI log. That mode replaces local root paths with a placeholder. Default JSON output keeps raw paths for scripts unless the redaction flag is passed. task show applies redaction to displayed task contract content. task list, artifacts, runs, show-run, intent, schemastore, and github import already use repo-relative paths or catalog metadata, and accept the flag for command consistency. agentloop review-context --json gives non-MCP agents one read-only local snapshot with status, active-task risk-note count, compact evidence map, context-budget estimates, gates, policies, artifacts, recent runs, latest ship evidence, and the next action. It omits the full changed-file list by default and points agents to agentloop context show evidence-map:current when they need file-level detail. agentloop upgrade-harness is read-only. It tells existing users which generated guidance files need manual review after a CLI upgrade. It does not merge templates or overwrite local edits. agentloop schemastore --json prints the catalog entry for agentloop.config.json. Use it when preparing a SchemaStore contribution; the CLI does not submit that contribution for you. schemastore --redact-paths is accepted for consistency; catalog values are not local filesystem paths, so output values do not change. agentloop github import --issue-json issue.json --pr-json pr.json imports explicit local GitHub metadata into .agentloop/github/context.json. review-context, prepare-pr, and maintainer-check use that local context when it exists. Missing metadata does not block the loop. github import --redact-paths is accepted for consistency; metadata paths are already repo-relative, so output values do not change. The CLI does not call GitHub APIs, read tokens, or run gh. agentloop release-proof checks post-release evidence for npm, GitHub Releases, GitHub Marketplace, GHCR, and MCP Registry. Use --only <channel> to re-check one channel. It can query public metadata or read captured JSON files that you pass explicitly. It does not publish, tag, upload, post comments, or read tokens. Shell completions include fixed values for task types, task-list status filters, status transitions, the done archive status, agent names, policy subcommands, completion shells, artifact types, badge sources, summarize and handoff formats, and release-proof --only channels.

More Docs

• Getting started
• CLI reference
• Upgrading existing repos
• Configuration
• Task contracts
• Research tasks
• AgentLoop Context Contract
• Evidence map and resume packs
• Verification reports
• Status and next action
• Gate checks
• HTML reports
• PR summaries and handoffs
• Policies
• SchemaStore support
• GitHub metadata import
• Real-repo trials
• Release proof
• MCP server
• Template and harness upgrades
• GitHub Actions and Marketplace Action
• Distribution channels
• Comparison

Examples

See examples/ for sample generated harnesses and CI recipes:

• End-to-end workflow
• Bugfix PR
• Next.js app
• Node API
• Python service
• Docs-only repo
• Empty repo
• GitHub Actions
• GitLab CI
• Buildkite
• Security review
• Dependency upgrade

See docs/stack-recipes.md for Next.js, React/Vite, Remix, SvelteKit, Node API, Django, FastAPI, Python, docs-only, empty-repo, and monorepo verification recipes.

Compared With Plain AGENTS.md

AGENTS.md gives instructions. AgentLoopKit adds process artifacts: task contracts, gates, policies, verification reports, PR summaries, and rollback notes.

The result is easier to review because the work produces evidence, not only code.

Compared With Prompt Collections

Prompt collections help an agent respond better. AgentLoopKit gives the repository a standing workflow that compatible agents can read and follow.

Compared With Superpowers

Superpowers is a broader agent methodology and skills system. AgentLoopKit focuses on repo-level engineering loops, task contracts, verification evidence, and handoff artifacts.

They can be complementary. You can use AgentLoopKit alongside Superpowers. AgentLoopKit has no affiliation with Superpowers.

Contributing

Contributor setup:

git clone https://github.com/abhiyoheswaran1/agentloopkit
cd agentloopkit
pnpm install
pnpm test
pnpm build

See CONTRIBUTING.md for local setup and PR expectations.

License

MIT.