Claude Code plugin that delegates to isolated CLIs and verifies before

The first draft repeats the same facts across sections and runs long. I'm tightening each section so the names and numbers stay, and each paragraph lands in range.PyModel's Claude Architect is a public-beta Claude Code plugin. PyModel's Claude Architect repository says Claude writes the spec, judges the evidence, and reports what landed. A fresh-context Producer, Codex, OpenCode, Pi, Pythinker, Antigravity CLI, or a headless Claude Code session, starts clean in an isolated Git worktree and returns a frozen, hash-anchored candidate. Claude reviews it against independent verification evidence before a single byte can reach the checkout.

Developers who run Claude Code in a Git repository and want a reviewed handoff are the audience. This covers guarantees, guarded integration, Node.js 22, Git 2.40, lane overrides, decision authority, skills, agents, MCP tools.

What shipped in Claude Code plugin that delegates to isolated

The install name is architect@claude-architect, from the PyModel/claude-architect marketplace. Claude writes the versioned spec, judges the evidence, and reports what landed, and one Producer implements the change. Three guarantees are enforced in host code: isolation, evidence over claims, and separated authority. Isolation uses a detached worktree, a sanitized environment, an explicit write allowlist, and OS sandboxing where certified. Out-of-scope changes are rejected at freeze time. The project is a public beta and says not to use Claude Architect unattended for production, destructive, or security-sensitive work.

A versioned spec moves through an isolated-worktree Producer, a frozen candidate, independent verification, adversarial review, and the configured decision authority. Accept leads to guarded integration, and reject or revise discards the candidate or starts a fresh attempt. Review, decision, and hash-gated integration are separate, and integration stages the reviewed tree without committing it. The central rule: agent output is an untrusted candidate, implementers cannot approve their own work, and only an independently verified pipeline candidate with commit-bound gate clearance can be accepted without a human.

What improved in Claude Code plugin that delegates to isolated

A Producer saying "tests pass" is never accepted. The runtime reruns the authorized verification commands in a clean worktree and records the real output. Candidate bytes are frozen and identified by hashes before that rerun. Reviewers work in a fresh context, read-only roles lack mutation tools, and the runtime rejects nested delegation, scope escapes, changed bases, mismatched anchors or trees, and unaccepted candidates. /architect:delegate keeps an isolated worktree, a frozen Candidate Artifact, independent verification, and guarded integration. /architect:codex runs codex exec on the current checkout.

OpenCode and Antigravity CLI accept model, thinking, variant, and effort. Pythinker accepts provider and model, with no reasoning override. The Claude Code lane accepts Opus or Sonnet plus reasoning effort. Codex accepts model and reasoning effort. Pi accepts thinking level only, and a model override fails the lane instead of substituting one. Unavailable lanes are reported rather than silently substituting another agent. Availability depends on host OS, CLI version, authentication, the requested lane, and proven execution capabilities. Non-trivial work uses the fresh-context review pipeline on the patch, findings, and verification output.

What you gain from Claude Code plugin that delegates to isolated

Producers authenticate as codex, opencode, pi, pythinker, agy, or claude. candidate-reviewer is a read-only Opus reviewer for one frozen candidate. Through reviewCandidate it reads the exact bytes, returns two verdicts and a recommendation, and never decides or integrates. advisor is a strictly read-only commitment-boundary advisor using bounded, redacted gitStatus, gitDiff, gitLog, and gitChangedFiles. delegation-lane, described in the delegate skill, is a courier limited to delegate and delegatePipeline. It cannot read the repository, run commands, review, decide, or integrate. Different repositories run in parallel, and the same repository is serialized by the runtime's repository lock.

/architect:delegate builds a versioned spec and drives delegation, review, decision, and guarded integration. /architect:subagent-driven-delegation runs a multi-task plan on the Superpowers subagent-driven-development loop, with a verified Producer on every task. validateDelegationSpec checks a spec without starting a Producer and returns its canonical correlation digest. delegate runs one isolated, independently verified attempt. delegatePipeline runs the fresh-context implement, review, and repair pipeline. reviewCandidate returns the frozen patch and verification evidence. decideCandidate records accepted, rejected, or revision-requested. integrateCandidate applies an accepted hash-matched candidate under safety guards. doctor reports runtime, Git, platform, and Producer diagnostics.

How to get Claude Code plugin that delegates to isolated

Claude Code requires Node.js 22 or newer, and the runtime requires Git 2.40 or newer so review diffs can read attributes from the trusted base with --attr-source. doctor reports git-too-old below 2.40. Install adds the PyModel/claude-architect marketplace, installs architect@claude-architect, lists plugins with --json, and restarts Claude Code after installing or updating. At least one of codex, opencode, pi, pythinker, agy, or claude is installed and authenticated. Claude Code is opened in a Git repository. If no Producer is named, /architect:delegate asks for one.

claude plugin marketplace add PyModel/claude-architect
claude plugin install architect@claude-architect
claude plugin list --json

The /architect:delegate quick start tells Codex to add rate limiting to the public API, run the tests, and show the independently reviewed candidate before integration. /architect:codex runs codex exec to review the checkout with gpt-6.1-sol at high reasoning. CLAUDE_ARCHITECT_DECISION_AUTHORITY=human requires confirmation for every decision, and an unrecognized value fails closed to human with a warning. CLAUDE_ARCHITECT_JEV=on with a TYPESAFE_API_KEY adds a Jev screen that can send an autonomous acceptance to a human, never the reverse. Omitted model fields use the Producer CLI's configured default.

/architect:delegate Use Codex to add rate limiting to the public API, run the tests, and show me the independently reviewed candidate before integration.

/architect:codex Review this checkout with gpt-6.1-sol at high reasoning.

CLAUDE_ARCHITECT_DECISION_AUTHORITY=human

CLAUDE_ARCHITECT_JEV=on

What to watch after Claude Code plugin that delegates to isolated

The public-beta notice says to review the complete candidate and the verification evidence before integration. Verification reduces risk and does not establish that a change is safe for a particular deployment. docs/SECURITY_MODEL.md and docs/THREAT_MODEL.md state that limit. A human is required for gate-refused, review-incomplete, malformed, commit-mismatched, human-required, unverified, and unreadable cases, and for every non-accept verdict. Independent verification decides what may be accepted at all. Integration refuses unknown provenance and aborts on a moved HEAD, a dirty tree, or a hash that does not match the reviewed artifact. Every decision records its provenance.