Skip to Content
ConfigurationTheory of Mind

Theory of Mind (tom_features)

Configures Navigator’s Theory of Mind layer — verification checkpoints, bilateral preference modeling, and collaboration quality detection. Based on Riedl & Weidmann (2025): better mutual modeling between human and AI predicts collaborative ability, so Navigator makes that modeling explicit.

Config block

In .agent/.nav-config.json:

{
  "tom_features": {
    "verification_checkpoints": true,
    "confirmation_threshold": "high-stakes",
    "profile_enabled": true,
    "diagnose_enabled": true,
    "belief_anchors": false
  }
}

Keys

  • verification_checkpoints (default true) — Show understanding before generating code on high-stakes skills. Claude restates the intended change so you can correct it before any code is written.
  • confirmation_threshold (default "high-stakes") — When checkpoints fire. "high-stakes" limits them to the high-stakes skill set; "always" confirms on every generation; "never" disables confirmation.
  • profile_enabled (default true) — Enable bilateral modeling via nav-profile. Claude learns and reuses your preferences across sessions.
  • diagnose_enabled (default true) — Enable quality detection via nav-diagnose. Detects when collaboration alignment degrades and prompts re-anchoring.
  • belief_anchors (default false) — Emit explicit assumption declarations (known / assumed / unknown) before generation. Opt-in; adds a short preamble in exchange for surfacing hidden assumptions early.

High-stakes checkpoint set

Verification checkpoints apply only to skills where a wrong assumption is expensive:

  • backend-endpoint
  • frontend-component
  • nav-task
  • database-migration

Simple operations skip the checkpoint, so the overhead lands only where it pays off.

Bilateral modeling (nav-profile)

With profile_enabled, nav-profile maintains a persistent profile of your communication style, preferred frameworks, verbosity, and testing approach. It auto-learns from corrections — when you correct the same thing, the pattern is captured so you don’t repeat it. Update explicitly with “Remember I prefer concise explanations”.

Quality detection (nav-diagnose)

With diagnose_enabled, nav-diagnose watches for collaboration quality drops: 2+ corrections on the same topic, references to non-existent files, context bleed between unrelated tasks, or goal drift. When triggered it pauses to re-anchor on the original intent rather than continuing off-target. Invoke manually with “something seems off”.

Metrics

ToM overhead and its effect on session quality are instrumented with OpenTelemetry. Run /nav:stats to see checkpoint frequency and correction counts for your own sessions, rather than relying on quoted figures.

Example

{
  "tom_features": {
    "verification_checkpoints": true,
    "confirmation_threshold": "always",
    "profile_enabled": true,
    "diagnose_enabled": true,
    "belief_anchors": true
  }
}

A high-assurance setup: confirm understanding on every generation and declare belief anchors before writing code.

OverviewLoop Mode