intertrust

intertrust

Agent trust scoring for Claude Code. Tracks which review agents produce useful findings and which waste tokens.

What this does

When interflux dispatches review agents, some consistently produce findings you act on — and some produce noise you dismiss. Intertrust closes the feedback loop: it records each accept/dismiss decision, computes a trust score per agent, and feeds that score back into dispatch priority so the good agents run first and the noisy ones get deprioritized.

The scoring algorithm uses severity-weighted time decay: a P0 finding accepted yesterday counts more than a P3 finding dismissed last month. Scores blend project-specific data with global data, so a new project inherits the agent's cross-project reputation until enough local data accumulates.

Intertrust was extracted from the interspect profiler to maintain single-responsibility: interspect handles evidence collection and routing overrides; intertrust handles reputation and trust.

Install

First, add the interagency marketplace (one-time setup):

/plugin marketplace add mistakeknot/interagency-marketplace

Then install:

/plugin install intertrust

Usage

Check which agents are earning trust and which are candidates for suppression:

/trust-status

AGENT                PROJECT              TRUST ACCEPTED  DISCARD    REVIEWS
fd-safety            my-project            0.92        18        2         20
fd-correctness       my-project            0.85        12        3         15
fd-game-design       my-project            0.15         1       12         13 <!>

Agents with trust < 0.30 are flagged with <!> as suppression candidates. These agents consistently produce findings that nobody acts on.

For a specific agent:

/trust-status fd-safety

How trust scores work

Score range: 0.05 (floor) to 1.0 (ceiling).

Inputs: Every time you resolve a review finding (via /clavain:resolve), the outcome is recorded:

• Accepted — you fixed the code to address the finding
• Discarded — you dismissed it as irrelevant or incorrect

Severity weighting: A P0 finding counts 4x, P1 counts 2x, P2 counts 1x, P3 counts 0.5x. Catching a real security issue (P0, accepted) boosts trust much more than flagging a style nit (P3, accepted).

Time decay: Half-life of ~30 days. Recent outcomes matter more than old ones. An agent that improved its prompts last week shouldn't be penalized for noise it generated two months ago.

Project/global blending: New projects inherit the agent's global reputation until enough local data accumulates (blend weight reaches 1.0 at 20 local reviews).

Integration: interflux multiplies each agent's triage score by its trust score at dispatch time. High-trust agents get dispatched first. Low-trust agents may not get dispatched at all if the token budget is tight.

Disabling

Trust scoring is progressive enhancement — it never blocks workflows. If intertrust is not installed, all agents get a neutral trust score of 1.0.

Architecture

Trust data lives in the shared .interspect/interspect.db SQLite database (the trust_feedback table). The library is self-contained with no dependency on the interspect plugin — it creates its own table if needed.

intertrust/
├── .claude-plugin/plugin.json    # Plugin manifest
├── hooks/
│   └── lib-trust.sh              # Trust scoring library (233 lines)
├── commands/
│   └── trust-status.md           # /trust-status command
└── tests/
    └── test_trust_scoring.sh     # 11 tests

Related

• interspect — agent profiler (evidence, routing, canaries)
• interflux — multi-agent review engine (the primary trust consumer)
• Clavain — the orchestrator whose resolve command feeds trust data