cross-model-verify

Cross-Model Verify

Use this skill when a task's contract has cross_model: required and you need to collect panelist verdicts before running the deterministic gate.

Never run this skill in CI. Collection is local-only and opt-in (design.md §6).

Prerequisites

• Task has Cross-Model: enabled in specs/<feature>/tasks.md (or a valid SDD_SUDO token exists).
• prepare-panelist-input.{sh,ps1} has been run and produced a sanitized bundle at specs/<feature>/verification/<task>.panelist-input.txt.
• At least one non-Anthropic panelist CLI is available (codex or gemini), OR you will provide a manual verdict JSON after running this skill.

Blind & Parallel Isolation Rules

Panelists MUST be invoked blind. Before calling any panelist:

1. Do NOT share any other panelist's verdict, findings, or model with the next panelist.
2. Do NOT share the primary evaluator's verdict, confidence, or reasoning.
3. Do NOT pass prior review-ticket content, Implementation Complete reports, or any specs/<feature>/verification/T-NNN.cross-model.json to panelists.
4. Each panelist receives ONLY the sanitized input bundle (T-NNN.panelist-input.txt) — nothing else from the repository context.

These rules are enforced structurally: Claude panelist runs via Agent tool in a fresh context; GPT/Gemini panelists run via isolated CLI runners with key isolation (see run-panelist-gpt.{sh,ps1} and run-panelist-gemini.{sh,ps1}).

Steps

Step 1 — Consent + Sanitize

Run prepare-panelist-input if not already done:

# bash
bash plugins/sdd-quality-loop/scripts/prepare-panelist-input.sh \
  --task <T-NNN> --feature <feature> \
  --input specs/<feature>/ \
  --spec-root specs

# pwsh
pwsh plugins/sdd-quality-loop/scripts/prepare-panelist-input.ps1 \
  --task <T-NNN> --feature <feature> \
  --input specs/<feature>/ \
  --spec-root specs

This will:

• Fail closed if Cross-Model: enabled is absent and no SDD_SUDO token exists.
• Strip secrets, absolute paths, and private URLs.
• Write specs/<feature>/verification/<T-NNN>.panelist-input.txt.
• Print the input_digest (64-hex SHA-256) to stdout.

Record the input_digest for the --expect-digest gate check later.

Step 2 — Detect available panelists

bash plugins/sdd-quality-loop/scripts/detect-panel.sh
# or
pwsh plugins/sdd-quality-loop/scripts/detect-panel.ps1

• Exit 0 + slug list → proceed.
• Exit 1 (no non-Anthropic CLIs) → warn user; they must provide a manual verdict JSON for at least one non-Anthropic vendor, or install a CLI.

Step 3 — Invoke panelists blind in parallel

Run ALL panelists simultaneously. Do not wait for one before starting another.

Claude panelist (Anthropic vendor — always included):

Use the Agent tool to invoke sdd-panelist-gpt or sdd-panelist-gemini roles — actually, Claude itself is the Anthropic panelist. Invoke a fresh Agent with ONLY the sanitized bundle as input. The agent role is in plugins/sdd-quality-loop/agents/panelist-gpt.md (for GPT) or panelist-gemini.md (for Gemini-style structured review).

For the Anthropic (Claude) panelist: create a new Agent tool call with the sanitized bundle path and the instruction to return a cross-model-verdict/v1 JSON, writing it to specs/<feature>/verification/<T-NNN>.panelist-anthropic.verdict.json.

GPT panelist (OpenAI vendor — if gpt slug in detect-panel output):

bash plugins/sdd-quality-loop/scripts/run-panelist-gpt.sh \
  --task <T-NNN> --feature <feature> \
  --input specs/<feature>/verification/<T-NNN>.panelist-input.txt \
  --spec-root specs

Gemini panelist (Google vendor — if gemini slug in detect-panel output):

bash plugins/sdd-quality-loop/scripts/run-panelist-gemini.sh \
  --task <T-NNN> --feature <feature> \
  --input specs/<feature>/verification/<T-NNN>.panelist-input.txt \
  --spec-root specs

Each runner writes its verdict to specs/<feature>/verification/<T-NNN>.panelist-<vendor>.verdict.json.

Step 4 — Verify verdict files present

After all panelists complete, confirm verdict files exist:

ls specs/<feature>/verification/<T-NNN>.panelist-*.verdict.json

Minimum required: one anthropic + one non-Anthropic (openai or google). If a runner exited non-zero (graceful degrade), that vendor is absent — the gate will fail diversity unless a manual verdict is provided.

Step 5 — Prompt user to run the gate

Tell the user to run check-cross-model to apply the deterministic consensus:

bash plugins/sdd-quality-loop/scripts/check-cross-model.sh \
  --task <T-NNN> --feature <feature> \
  --evaluator <PASS|NEEDS_WORK> \
  --expect-digest <input_digest_from_step1>

# or pwsh
pwsh plugins/sdd-quality-loop/scripts/check-cross-model.ps1 \
  --task <T-NNN> --feature <feature> \
  --evaluator <PASS|NEEDS_WORK> \
  --expect-digest <input_digest_from_step1>

The gate writes specs/<feature>/verification/<T-NNN>.cross-model.json and exits 0 (PASS), 1 (FAIL/NEEDS_HUMAN), or 2 (tool error).

Graceful Degrade

If detect-panel exits 1 (no non-Anthropic CLIs):

• Inform the user: "No non-Anthropic panelist CLIs found. The cross-model gate will fail diversity unless you provide a manual verdict JSON."
• Instruct them to manually create specs/<feature>/verification/<T-NNN>.panelist-openai.verdict.json (or google) following the cross-model-verdict/v1 schema in references/cross-model-verification-policy.md.

Do NOT crash or block — the user may still proceed to manually provide verdicts.

Security Reminders

• The sanitized bundle in panelist-input.txt must be the ONLY input to each panelist. Do not add repo context, agent memory, or prior verdicts.
• SDD_EVIDENCE_KEY and SDD_SUDO_KEY are never passed to panelists.
• The collection layer never runs in CI (enforced by AC-007).