Cross-Check v2.1

Install: clawhub install cross-check

Verify assumptions in your responses. Opt-in — the agent suggests verification, you decide.

Capabilities Used

sessions_spawn — For 2-model verification mode (optional). Requires a second configured model. Only used when user explicitly requests "cross-check 2-model".
HEARTBEAT.md — Reads (never writes) to check if user has enabled auto-suggestions.

Language

Detect from the user's message language. Default: English.

How It Works

Default: Suggest, Don't Auto-Run

When the agent detects a complex response (3+ assumptions), it appends a one-line suggestion:

💡 Cross-Check available — reply "cross-check" to verify these assumptions.

The user chooses whether to activate. No silent auto-invocation.

User Activates

| Command | Action | |---------|--------| | "cross-check" / "sjekk dette" | Lite mode (2 rounds) | | "cross-check deep" | Deep mode (3 rounds or 2-model) | | "cross-check 2-model" | 2-model mode (requires sessions_spawn + second model) | | "cross-check off" | Disable suggestions for this session |

Opt-In Auto-Suggestions via HEARTBEAT

If the user adds the following to their HEARTBEAT.md:

## Cross-Check
- auto-suggest: true

...then the agent will suggest Cross-Check when it detects 3+ assumptions, without the user needing to trigger it first. This is still a suggestion — the user must reply "cross-check" to actually run it.

Three Output Levels

Default — Confidence Note

For responses with 1-2 assumptions, append:

Confidence: [High / Medium / Low]
Key assumption: [the main assumption]

Lite — 2 Rounds (same model)

Round 1 "The Analyst": Solve fully, extract assumptions. Round 2 "The Challenger": Solve from scratch, different angles.

Output (max 8 lines):

Cross-Check (Lite):
  Agreement: [what both agreed on]
  Difference: [where they disagreed]
  Blind spot: [thing neither considered]
  Confidence: [High / Medium / Low]

Deep — 3 Rounds or 2-Model

Option A: Reinforced (same model, 3 rounds) Round 3 "The Synthesizer": Both answers visible, finds consensus/divergence/blind spots. Includes pre-mortem.

Option B: Cross-Check (second model) Uses sessions_spawn to run a verifier sub-agent. Requires a second configured model.

Step 1: Primary solves, extracts assumptions
Step 2: Verifier challenges each assumption from 4 perspectives (Skeptic, Expert, Beneficiary, Contrarian)
Step 3: Primary integrates challenges

Output (max 15 lines):

Cross-Check (Deep):
  Mode: [Reinforced / Cross-Check]
  Consensus: [findings all rounds agree on]
  Divergence: [where rounds disagreed + resolution]
  Blind spots: [things none considered]
  Assumptions:
    - [assumption]: [confidence] — [confirmed/challenged/revised]
  Confidence: [High / Medium / Low]

Assumption Tracking

Every round tracks: core assumptions, confidence (High/Medium/Low), unknowns, biases.

Guidelines for Agent

Suggest, don't auto-run — show "Cross-Check available" line, let user decide
Respect "cross-check off" — disable suggestions for the session
Check HEARTBEAT.md — if auto-suggest is enabled, suggest proactively
Compact output — max 8 lines lite, 15 deep
Never modify files — reads HEARTBEAT.md only
2-model is optional — only mention if user asks or has multiple models
Cost awareness — lite = ~2x tokens, deep = ~3x tokens

Privacy and Safety

Session-only — nothing persisted
No personal data written anywhere
Verifier receives only problem context + assumptions
No file writes, no web searches unless user requests
Uses only OpenClaw's configured providers via sessions_spawn

What This Skill Does NOT Do

Does NOT auto-run verification without user opt-in
Does NOT modify any files
Does NOT replace the primary model
Does NOT persist anything
Does NOT send raw user data externally

More by TommoT2

setup-doctor — Diagnose and fix OpenClaw setup issues
context-brief — Persistent context survival across sessions
tommo-skill-guard — Security scanner for installed skills
locale-dates — Format dates/times for any locale

Install the full suite:

clawhub install setup-doctor context-brief tommo-skill-guard locale-dates