globus-mcp: Skill for Claude Code

globus-review — adversarial QA (clean context)

When to Use

Invoke /globus-review when a branch's TECH.md is fully built (status: in_review). Best run in a fresh session — but the real guarantee comes from delegating scrutiny to freshly-spawned subagents with curated inputs, so bias is removed even if this session is not clean. The reviewer grades the diff against the locked GOAL.md and the AGENTS.md invariants, by executed command, not opinion.

Operating manual: .agents/factory/review-rubric.md and .agents/factory/invariants.md. Read them before delegating.

Harness portability. Runs on any harness — see factory/portability.md. Fallbacks: run the Current state commands yourself if not auto-injected; ask in plain text and STOP if AskUserQuestion is unavailable; if subagents are unavailable, perform the correctness pass yourself in a clean context (you lose delegated blindness — compensate with executed evidence, per the rubric); and skip ReportFindings (REVIEW.md is the durable record). The production-safety rules are not portable-away — see the rubric.

User Instructions

Additional instructions provided with the invocation: $ARGUMENTS

Current state (injected at load)

• Branch: !git branch --show-current
• Base: main (feature/fix branch base; confirm from base: in TECH.md during Step 1).
• Diffstat vs main: !git diff --stat main...HEAD 2>/dev/null | tail -n 20

Argument Parsing

• status → report the current review verdict from TECH.md and any existing REVIEW.md; no work.
• debate → run the two-independent-reviewer variant (for high-risk / coupled-core diffs).
• completeness → also run the separate completeness sub-pass (may see TECH.md).

Safety Principles

• Blindness is the point. The correctness reviewer subagent is given GOAL.md, the diff, the runnable repo, invariants.md, and review-rubric.md — and is explicitly told NOT to read PLAN.md, TECH.md, research/, or META.md (the last leaks author intent / harness notes, same reason as PLAN/TECH). Only this skill (the orchestrator) reads TECH.md, and only for the base/slug/kind metadata — it must not pass PLAN/TECH content into the reviewer prompt. The diff must be blind too: those artifacts are committed on the branch, so a plain git diff {base}...HEAD hands the reviewer PLAN/TECH/research (and any prior cycle's REVIEW.md) as added hunks — the ':(exclude)spec/' pathspec below is load-bearing, not cosmetic.
• External verification is the spine — and it never touches production. Every finding must cite an executed command (uv run pytest, a client round-trip through .agents/factory/bin/mcp_probe.py, a throwaway driver that monkeypatches globus_mcp.cli.globus). No assertion-only findings. Equally: no finding may be produced by calling a mutating tool or using real credentials. Drives run under .agents/factory/bin/temp_home.sh. "Not reproducible without live credentials" is PLAUSIBLE, never CONFIRMED, and never a reason to reach for credentials.
• Refute before reporting. Try to disprove each candidate; classify CONFIRMED (reproduced) vs PLAUSIBLE (needs human triage). Default to dropping when uncertain.
• Scope is narrow: correctness bugs, GOAL R-ID gaps, AGENTS.md invariant violations (auto-CRITICAL), and scope creep (changes mapping to no R-ID). No style nits, no speculative hardening — a gap-hunting reviewer manufactures gaps.
• Read-only session. This skill makes no source edits; it writes REVIEW.md and updates the TECH.md review block via set_phase.py.
• Mandatory human gate when any CONFIRMED finding touches the high-blast-radius core (src/globus_mcp/tools/__init__.py, cli.py, compute.py, server.py, __init__.py) or a safety invariant (the consent contract, the shell=False subprocess boundary, the stdio-channel rule, the mutating-tool list, credential handling) — regardless of auto-loop.
• Bounded loop: ≤ 2–3 review↔build cycles; escalate to the human on non-convergence.