Kungfu
Keep one body of work moving across agents, sessions, and failures.
The README says the installer uses no sudo and does not edit shell profiles, while Kungfu never stages, commits, or pushes for the user. Workflows declare narrow permissions, pin several Actions to commits, and disable persisted checkout credentials. Independent review, settlement, and single-controller behavior provide evidence of confirmation and authority separation, and the project-local state flow is partly described. Deductions apply because the installer source, a complete runtime permission inventory, telemetry/network disclosures, and credential-storage implementation are absent. Sensitive-data evidence is primarily a reporting policy rather than operational controls. Durable Work supports recovery, but the supplied files do not show complete uninstall, data rollback, or external-effect inventories. Repository attribution and contact routes are clear, while publisher identity remains unverified.
The README's Project, Work, Attempt, single-writer, and independent-settlement model agrees with terminal tests for control ownership, blocked input, recovery presentation, and delivery outcomes, providing strong internal consistency. Tests expose specific actionable states, and workflows use strict validation with explicit error paths. Dependency availability is deducted because normal use still requires external Agents such as Codex, Claude, OpenCode, or Amp; the product is Alpha; and the supplied evidence omits the complete lockfiles and support-matrix contents.
The evidence covers first-run onboarding, recovery, multi-step approval, native Agent consoles, TUI/GUI, CLI/API use, and contributor scenarios. It clearly bounds the Mock Agent, real-Agent requirement, sidecar interfaces, completion authority, and the non-certifying nature of demonstrations. Commands, scenario variables, and controller conditions are reasonably precise. Deductions apply because Windows, higher-assurance installation, compatibility, and provider-specific behavior are only referenced through unavailable documents, and complete command implementations are not included.
The README has strong progressive information architecture, defines its core terminology, and separates installation, concepts, limits, security, upgrades, and contribution routes. Installation examples, PATH guidance, Alpha labeling, known-limit routes, and the full Apache-2.0 license are explicit. Naming is stable between documentation and tests. Deductions apply because there is no conventional FAQ, some examples depend on omitted documents, release and migration history is represented only by links or script names rather than an included changelog, and maintenance channels do not identify accountable maintainers, response commitments, or a verified publishing entity.
The product's output is directly usable: it retains objectives, progress, evidence, next actions, and completion state, with cross-Agent continuity, recovery, writer-conflict prevention, and readable UI states. This supplies clear marginal value beyond copying chat context. Cost-benefit is deducted because claims about choosing cheaper Agents and seeing value within minutes are not accompanied by quantified operating cost, migration effort, storage growth, or comparative evidence.
Core claims are separated into three bounded proofs with exact artifact references, while README statements, tests, and workflows corroborate continuity, control ownership, failure states, permissions, and source binding. The text explicitly separates deterministic Mock scenarios from real providers, artifact demonstrations from production certification, and Agent output from independent settlement. Full scores reflect strong static traceability and inference discipline only; they do not imply that this review executed or independently validated the proofs.
- The project is explicitly Alpha; do not treat demonstration artifacts or declared badges as production certification.
- Real workflows still depend on external Agents. Review their credential handling, network transfers, costs, and provider-compliance boundaries before use.
- Pin a version and inspect the installer before using the curl-to-shell command; install.sh was not included in the supplied evidence.
- Do not delete or wholesale commit .kungfu/. Inspect the workspaceGit policy and validate backup, uninstall, and data-recovery procedures for important projects.
- The publisher is not verified by the FollowAgents enterprise registry. This means identity is unknown, not suspicious.
What does this agent do, and when should you use it?
Kungfu is a durable Work layer for multi-agent development workflows, publicly available as a v4 Alpha. Instead of transferring chat history, it stores the objective, progress, evidence, next action, and completion state outside any individual Agent Session. Its model consists of Projects, Works, and Attempts: a Project groups related activity, a Work holds the current truth, and an Attempt records what one agent tried, including failures. The same Work can be accessed through native agent consoles, the `kungfu` CLI, TUI, GUI, and APIs, with documented launch commands for Codex, Claude, OpenCode, and Amp. Runtime state lives in a project-local `.kungfu/` workspace, while Git staging, commits, and pushes remain under the user's control. It is a fit for teams that need cross-agent continuity, recoverable execution, and independent completion review, but its Alpha status and reliance on a supported external agent for regular onboarding limit production readiness.
From a project directory, a user creates and executes Work with kungfu run <agent> "<task>", or asks an existing agent to run kungfu agent brief and guide initial Project and Work creation. Kungfu writes the Work objective, progress, evidence, next action, and completion state to the project-local .kungfu/ workspace. Each execution becomes an Attempt, so a new process can resume the same Work after a disconnect or crash without erasing earlier failures. kungfu agent map --json exposes workspace information and the workspaceGit policy for checking what may be staged. kungfu agent-work-lab provides an open → watch/tour → try → test → report verification journey, while the credential-free Mock Agent supplies deterministic recovery-story and multi-step scenarios. An executing agent may produce a candidate result and evidence, but independent review and Kungfu settlement control completion; Kungfu also blocks a second writer when another live agent already owns the Work.
- A developer who switches among Codex, Claude, OpenCode, or Amp and wants to continue one task without copying chat context.
- A team evaluating recovery behavior can run the deterministic Mock Agent
recovery-storyscenario through disconnect, crash, and resumed delivery. - An engineering organization that requires separation of duties can keep candidate production, independent review, and settlement distinct.
- A team maintaining long-running or multi-stage changes can preserve Attempts, current evidence, remaining work, and the next action.
- An evaluator without model-provider credentials can exercise the built-in recovery and multi-step workflows before connecting a real agent.
- A user managing Work across several projects can optionally open the Kungfu TUI or GUI as a global sidecar view.
What are this agent's strengths and limitations?
- Persists structured Work state rather than chat, retaining the objective, progress, evidence, and next action across fresh agent sessions.
- The Attempt model preserves disconnects and crashes while allowing a new process to resume the same Work.
- Separates execution from independent review and settlement, so the producing agent cannot approve its own Work.
- Documents operation with Codex, Claude, OpenCode, Amp, and custom execution surfaces through CLI, native consoles, TUI, GUI, and APIs.
- Includes deterministic Mock Agent scenarios that test recovery and multi-step flows without provider credentials.
- Prevents two live agents from silently diverging by stopping a second writer when a Work already has an owner.
- Kungfu v4 is explicitly an Alpha prerelease, not a stable or generally available release.
- Regular onboarding and independent review still require a supported real agent, so the Mock Agent is not a complete zero-external-agent path.
- Adoption adds a project-local
.kungfu/workspace whoseworkspaceGitpolicy must be understood before staging files. - The demonstrations establish bounded artifact behavior only; they are not production certification, provider rankings, or authority to complete real Work.
- A direct installation command is supplied only for macOS and Linux; Windows users must consult separate installation guidance.
- The selected agent must be able to run local commands, and Kungfu does not automate Git staging, commits, or pushes.
How do you install or deploy this agent?
On macOS or Linux, run curl -fsSL https://kungfu.tech/install.sh | sh. The reviewed per-user installer does not use sudo or edit the shell profile; follow the exact PATH step printed after installation. The source points Windows users, higher-assurance installations, and explicit version pinning to the installation guide but does not include their copyable commands here. Downloading the installer requires network access. The built-in Mock Agent scenarios require no provider credentials.
How do you use this agent?
Enter a project directory and run KUNGFU_MOCK_AGENT_SCENARIO=recovery-story kungfu to watch one Work survive three Attempts: disconnect, crash, and recovered delivery. For a shorter check, run KUNGFU_MOCK_AGENT_SCENARIO=multi-step kungfu. To use a real agent, run kungfu run codex; replace codex with claude, opencode, or amp as appropriate. Create a Work directly with kungfu run codex "Prepare the release notes". To remain in an existing agent interface, have that agent run kungfu agent brief and guide Project and Work creation. When .kungfu/ appears, do not delete it or add the entire directory to Git; run kungfu agent map --json and follow its workspaceGit policy before staging.
How does this agent compare with similar options?
Compared with manually copying chats and re-explaining decisions when moving among Codex, Claude, OpenCode, or Amp, Kungfu keeps the Work objective, evidence, progress, and Attempt history outside those sessions. It complements rather than replaces these agents, and its TUI and GUI are optional sidecar views rather than substitutes for native agent consoles.
FAQ
Do I need OpenAI or Anthropic credentials to try it?
recovery-story and multi-step Mock Agent scenarios. Regular onboarding and independent review still require a supported real agent.Will Kungfu commit or push repository changes?
kungfu agent map --json and follow the reported workspaceGit policy before staging .kungfu/ content.