vetmanager-mcp: Skill for Codex

Super Review

Use this skill when the user asks for super-review, deep review, multi-agent review, review arbitration, Spark scout review, or cross-model review in this repository.

Contract

Use the same review protocol as .claude/commands/super-review.md, but adapt runtime-specific steps to Codex tools:

• scopes: changed default, related, full, stage:N
• flags: --no-spark, --no-arbitration
• output report: artifacts/review/{YYYY-MM-DD}-{scope_slug}-stage-{N}.md
• dismissed memory: append to artifacts/review/inadequate-findings-index.md
• findings schema: YAML items with severity, reviewer, category, file, lines, problem, why_it_matters, suggested_fix, confidence

If the Claude command and this skill differ, prefer it only for shared review policy: scopes, model matrix, finding schema, report format, adequacy rules, and cross-CLI arbitration contract. Do not copy Claude-only mechanics such as .claude/agents/*, Agent tool-calls, or Claude command syntax into Codex execution.

Model Routing

Spark/GPT side:

• gpt-5.3-codex-spark: scout/prepass only. Treat output as untrusted leads.
• gpt-5.4-mini or gpt-5.3-codex-spark: code/docs/tests candidate mining.
• gpt-5.4: observability and normal validation.
• gpt-5.5: security, architecture, product, hard disputes, and GPT-side aggregation if needed.

Claude side:

• opus: external arbitration from Codex runtime; security/architecture/product/aggregator quality bar.
• sonnet: fallback external arbitration and routine code/docs/tests checks.

Never let Spark decide final severity or merge verdict. Use the exact model name gpt-5.3-codex-spark. The shorter name gpt-5.3-spark is incomplete and must not be used.

Runtime Defaults And Fallbacks

Use these defaults when shelling out from Codex:

• Per reviewer/scout command timeout: 1200 seconds.
• External arbitration timeout: 900 seconds.
• First attempt sandbox: -s read-only.
• Classify failures before retrying:
  • Sandbox/runtime failure: process fails before meaningful file reads due to local execution errors such as bwrap: loopback: Failed RTM_NEWADDR, sandbox initialization, PTY/session startup, local timeout wrapper issues, or shell/env problems. Retry the same model once with the adjusted runtime settings described below. Do not switch models for these failures.
  • Model failure: stderr/stdout clearly indicates the requested model is unavailable, unknown, unsupported, denied for the account, capacity-limited, quota-limited, or otherwise rejected by the model provider. Only this class may use the documented fallback model.
  • Review-output failure: command ran and read files but returned empty output, malformed YAML, non-findings prose, or a meta runtime error. Do not switch models automatically unless the output itself identifies a model/provider failure; mark the role skipped_or_failed or rerun the same model once if the failure is obviously transient.
• If Codex CLI fails before reading files with sandbox/runtime errors such as bwrap: loopback: Failed RTM_NEWADDR, retry that exact pass once with -s danger-full-access and a prompt that explicitly says: Review only. Do not edit files. Do not run write commands.
• If and only if a model failure is identified, retry once with the documented fallback model and record the fallback in the report header/limitations.
• If a role returns empty output, non-YAML output, or only a meta runtime error, do not silently treat it as "no findings"; mark that role as skipped_or_failed in the report header and continue with available roles.
• Before finalizing, check that no codex exec / claude -p processes from this run are still alive.