Reviews plans, diffs, and decisions through a read-only cross-vendor model bridge.
npx skills add https://github.com/peregrinus879/eyragents --skill spar-peregrinus879
Please help me install this Agent Skill. Skill: spar Source: https://github.com/peregrinus879/eyragents/tree/main/agents/.agents/skills/spar Command: npx skills add https://github.com/peregrinus879/eyragents --skill spar-peregrinus879
spar-codex, OpenCode via spar-claude, with explicit reviewer-ID resumption and provenance reporting. - Evidence-backed artifacts: The review-brief script records HEAD, source/index fingerprints, gate commands and results, and scans artifacts for sensitive content before disclosure. - Structured adversarial loop: A reviewer charter demands grounded objections classified by impact, ending in VERDICT: CONVERGED or VERDICT: OPEN with blocking counts. - Use Case: Before presenting an implementation plan for approval, generate a review brief with gate evidence, run the cross-vendor bridge, address or rebut each objection, and present one ruling packet per open item to the human arbitrator. ## Quick Start Ask the agent to run a spar review of the current plan or staged diff using the review-brief script and the appropriate cross-vendor bridge, then report the reviewer's verdict and provenance.High-intent search queries and answers about installing and using this skill.
Generate an evidence artifact with the review-brief script, then invoke the spar-codex or spar-claude bridge with your request and artifact. The reviewer model returns grounded objections and a verdict of CONVERGED or OPEN.
Cross-model review sends an artifact to a different vendor's model so the implementer and reviewer are independent. Use it before plan approval or after implementation when an outside view could materially change the outcome; it is never mandatory.
No. A Codex session cannot launch spar-claude inside its own profile. It prepares the request and artifact, then asks the human operator to run the bridge and return the reply.
Artifacts are scanned for credential-shaped content, and repositories can opt out via git config spar.consent. Confirmed sensitive spans are replaced with a redaction marker only after a human ruling; content is never edited merely to pass the scan.
Gate execution refuses with exit 64 when MAKEFLAGS, MFLAGS, GNUMAKEFLAGS, or MAKEFILES are nonempty, since inherited flags could silently alter dry-run or error-control behavior. Worktree diffs also refuse configured clean filters or a present GIT_CONFIG.
Use the in-tool auditor when the review needs authorized context outside the repository or when an in-tool follow-up is more useful. Spar is for independent cross-vendor judgment; never present an in-tool audit as cross-vendor spar.
labs21-devTranslate vague product goals into verifiable specifications with user stories and acceptance criteria.
dpaola2Guide reviewers through pipeline checkpoint approvals and update artifacts with decisions.
VolinadValidate and polish documentation sets with cross-checks and fixes.
VMAFxRun a local LLM code review on a single file with vmaf-dev-llm.