skill_verify

npx skills add moveju112/skill_verify

A bidirectional Claude Code and Codex verification bundle. The active host owns the implementation while the other agent independently analyzes and verifies it. It also includes a host-neutral multi-angle unit-test pass shared by Claude and Codex.

Skill	Role
verify:crosscheck in Claude	Claude designs and writes the code; Codex independently analyzes and verifies completion.
verify in Codex	Ordinary requests stay local. An explicit Claude crosscheck makes Codex implement and Claude independently review.
unit-test in Claude or Codex	Approval-gated unit testing right after a code change, across 7 angles. Writes a report to <project>/test/.

Both skills work in English and Korean. Triggers are registered in both languages, and the output — chat replies and the test report — follows whichever language you write in.

Install

npx skills add moveju112/skill_verify

Installs the portable skills. This repository also contains a Codex-specific entrypoint at platforms/codex/verify for the reciprocal Claude-review flow.

The Codex crosscheck entrypoint is runtime-specific and is not installed from skills/. On a shared-agent host, register codex:verify as runtime-only and point the Codex verify skill link at platforms/codex/verify. Without a shared manager:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
ln -sfn "$PWD/platforms/codex/verify" "${CODEX_HOME:-$HOME/.codex}/skills/verify"

As a Claude Code plugin instead:

claude plugin marketplace add moveju112/skill_verify
claude plugin install verify@verify

Or from inside a Claude Code session:

/plugin marketplace add moveju112/skill_verify
/plugin install verify@verify

Requirements

• Claude-hosted crosscheck: Claude Code plus a logged-in Codex CLI.
• Codex-hosted crosscheck: Codex plus a logged-in Claude Code CLI and Linux Bubblewrap (bwrap). Calling Claude requires explicit task-scoped permission; generic verification never calls it.
• unit-test has no external dependency.

Usage

Say the trigger in either language, or call the slash command:

Intent	English	Korean	Command
Cross-verify with Codex	"crosscheck", "cross-check with codex", "get a codex review"	"크로스체크", "codex 교차검증", "코덱스랑 핑퐁"	/verify:crosscheck
Cross-verify with Claude from Codex	"cross-check with Claude", "ask Claude to review"	"클로드랑 크로스체크", "클로드 리뷰 받아"	$verify or natural language
Unit-test the last change	"unit test", "run the tests", "multi-angle tests"	"단위 테스트", "테스트 돌려", "다방면 테스트"	/unit-test, /verify:unit-test, or $unit-test

---

crosscheck

Claude and Codex analyze independently and blind, then exchange and merge findings. The agent running the user's session owns the code; the counterpart is an independent analyst and verifier, never an author.

Modes

Picked from the request; ambiguous requests default to full.

Mode	Phases	Example trigger
analyze	A	"merge both opinions" / "둘이 의견 취합해줘"
plan	A + B	"ping-pong it and just give me the plan" / "핑퐁해서 계획만 줘"
full	A + B + C + D	"cross-check it and do the work" / "교차검증하고 작업해" (default)
verify	D	"have codex check whether this is done" / "다 됐는지 codex 체크"

• A — parallel blind analysis. The reviewer call starts in the background from a pre-fixed prompt while the host analyzes independently. The exact task result is collected before comparison; dependent later phases remain sequential.
• B — merge into a plan. Disagreements are numbered and resolved.
• C — implementation. Claude only.
• D — completion check. A fresh Codex session judges whether the work meets the requirement, with the diff scope pinned to a baseline SHA. It is fired in the background by default; the run does not close before the result is collected.

Phases compose freely. Stop the run mid-way and you keep that phase's output.

Why the wrapper exists — token-leak safety

A naive codex exec call drags Codex's full reasoning trace, every file it read, and its raw logs back into Claude's context. This bundle routes every call through scripts/codex_ask.sh, which is the only supported entry point:

• Raw Codex output is confined to a log file; only the final message reaches stdout, truncated to 6,000 characters by default.
• File contents and diffs are never pasted into prompts. Codex reads the repository itself from a read-only sandbox; prompts carry paths and line ranges only.
• Logs may contain source read by Codex, so they are created with umask 077 (owner-only) and pruned after 14 days.
• resume accepts an explicit session UUID only. Resuming "the latest session" globally is deliberately unsupported — under parallel runs it attaches to someone else's thread.

Review discipline

• Findings arrive numbered and severity-tagged: #N [blocking|minor].
• Follow-ups reference numbers only ("#1 rebutted: <evidence>, #2 accepted and fixed") — never restate the finding. Cheaper, and it keeps the comparison exact.
• A Codex finding is a hypothesis until verified. Before acting on one, read the cited file:line and confirm the path is actually reachable. Measured on this bundle, 15 of 23 responses came back DISAGREE — and some of those had misread a project rule or the review scope.
• Round caps: 2 for analysis, 3 for planning, 2 for completion checks. On cap, Claude decides and records the dissent in the final report.
• Sessions are replaced, not extended forever: resume resends the whole thread each time (one measured confirmation round cost 4.5M input tokens). Past 3 rounds, or on a topic change, start a new session.

codex_ask.sh

codex_ask.sh check                                # preflight: login state, no tokens spent
codex_ask.sh new    [-C <repo>] "prompt"          # start a session
codex_ask.sh resume <session-uuid> "prompt"       # continue one (UUID only)
echo "long prompt" | codex_ask.sh new -C <repo> - # '-' reads the prompt from stdin

The first stdout line is SESSION: <id>; feed it to the next resume. Call it with a Bash timeout of 930000 ms — the script's own ceiling is 900 s, and a shorter caller timeout kills Codex from the outside and discards the round.

Env var	Default	Meaning
CROSSCHECK_MAX_CHARS	6000	stdout truncation limit
CROSSCHECK_TIMEOUT	900	per-call ceiling, seconds
CROSSCHECK_SANDBOX	read-only	Codex sandbox mode
CROSSCHECK_LOG_DAYS	14	log retention, days
CROSSCHECK_STATE_DIR	~/.cache/codex-crosscheck	log location
CODEX_MODEL	Codex default	model override
CODEX_EFFORT	Codex default	minimal｜low｜medium｜high

Status strings on stdout: CODEX_OK, CODEX_NOT_INSTALLED, CODEX_AUTH_ERROR, CODEX_QUOTA_ERROR, CODEX_TIMEOUT, CODEX_ERROR, CODEX_FORMAT_WARNING.

CODEX_FORMAT_WARNING means the response broke its output contract (missing VERDICT: first line, missing severity tag, non-sequential finding numbers, over 3,000 characters, or more than 7 bullets). Re-ask instead of guessing the verdict — except for a lone length warning, which is safe to accept.

claude_ask.sh (Codex host)

The reciprocal Codex entrypoint calls Claude only through platforms/codex/verify/scripts/claude_ask.sh:

claude_ask.sh check
CROSSCHECK_REMOTE_APPROVED=1 claude_ask.sh new -C <repo> "prompt"
CROSSCHECK_REMOTE_APPROVED=1 claude_ask.sh resume <session-uuid> -C <repo> "follow-up"

• check is local-only and verifies Claude login plus required commands.
• new and resume refuse to run without the task-scoped CROSSCHECK_REMOTE_APPROVED=1 gate.
• Bubblewrap mounts only the target repository (read-only), current Git evidence (read-only), and the current session's isolated home (writable). Host credentials, unrelated repositories, and older review sessions are absent.
• Claude receives only Read, Glob, and Grep; mutation, Bash, workflow, subagent, and external-tool entrypoints are disabled.
• The auth secret is passed through a one-use file descriptor. It is not mounted into the sandbox or inherited as a secret environment variable.

Env var	Default	Meaning
CROSSCHECK_REMOTE_APPROVED	unset	must be 1 for each approved Claude run
CROSSCHECK_BASE_SHA	HEAD	baseline for the precomputed review diff
CROSSCHECK_MAX_CHARS	6000	stdout truncation limit
CROSSCHECK_TIMEOUT	900	per-call ceiling, seconds
CROSSCHECK_LOG_DAYS	14	evidence, raw-log, and session retention
CROSSCHECK_STATE_DIR	~/.cache/claude-crosscheck	owner-only state root, outside the reviewed repo
CLAUDE_MODEL	Claude default	model override
CLAUDE_EFFORT	Claude default	effort override

Status strings on stdout: CLAUDE_OK, CLAUDE_NOT_INSTALLED, CLAUDE_PERMISSION_REQUIRED, CLAUDE_AUTH_ERROR, CLAUDE_QUOTA_ERROR, CLAUDE_TIMEOUT, CLAUDE_ERROR, CLAUDE_FORMAT_WARNING.

---

unit-test

Shared by Claude and Codex through the same host-neutral skill source. The active host runs the tests and never calls the counterpart or another model.

• Gate. After a code change, the skill offers to run the tests and proceeds only on approval. Invoking it directly counts as approval. External DB, network, runtime, and performance checks are excluded and require a separate verification step with its own permission.
• Scope. Changed functions from git diff HEAD, classified as pure logic / local fixtures / external verification. Unit testing stays local; DB, network, runtime, and performance checks are routed to a separate verification step.
• Angles. happy path, boundary, empty·null·missing, error path, ordering·determinism, before/after equivalence, side effects. Angles that do not apply are skipped and named as skipped in the report.
• Efficiency. Focused tests run first. One full-suite owner runs on the final stable change; an official build/release wrapper owns that gate when it already runs the suite, preventing duplicate full runs. Proven baseline failures are recorded once instead of rerun after every edit.
• Output. <project>/test/UNITTEST_<date>_<topic>.md, written in your language. Unit counts are validated against case rows, reports roll over at 300 lines or a changed objective, and external checks never enter unit totals. Script location is resolved in three steps: project rule docs → an existing test directory → create test/scripts/.

---

Development

bash tests/codex_ask.test.sh    # fake Codex CLI; no API calls
bash tests/claude_ask.test.sh   # fake Claude CLI; no API calls
bash tests/claude_sandbox.test.sh # real Bubblewrap boundary; no API calls
bash tests/review_git.test.sh   # temporary local Git repository
bash tests/unit_test_skill.test.sh # shared host-neutral unit-test contract

.claude-plugin/     plugin.json + marketplace.json
skills/crosscheck/  SKILL.md + scripts/codex_ask.sh
skills/unit-test/   SKILL.md + scripts/validate_report.py
platforms/codex/    reciprocal Codex host entrypoint + read-only wrappers
tests/              wrappers + sandbox/Git boundaries + shared unit-test contract

License

MIT — see LICENSE.