Credence

Credence helps AI agents choose a model and reasoning effort from a cold-start guide plus sparse evidence from real work.

It is local, dependency-free, and deliberately avoids scores, telemetry, and automatic reviews. The first automatic integration is for Codex; the CLI can store evidence for any model and effort pair.

Install

Requires Node.js 18+ and Python 3.10+.

CLI

npm install --global @relic-a/credence
credence read

Codex plugin

codex plugin marketplace add Relic-a/Credence
codex plugin add credence@credence

Start a new Codex thread after installation. Credence will consult its routing guide before an explicit subagent model or effort is selected.

Use

# Read saved evidence
credence read

# Record an observation directly
credence add --model "gpt-5.6-terra" --effort "high" \
  --observation "Strong verified result on a bounded repository migration."

# Inspect anonymous metadata after reviewing delegated Codex work
credence review-context

# Before review, omit a child known to be interrupted or otherwise invalid
credence exclude --agent-name "/root/<canonical-subagent-name>"

# Attribute an observation to a reviewed Codex subagent
credence add --agent-id "<subagent-uuid>" \
  --observation "Unexpectedly strong result with complete test coverage."

Credence suggests an exact merge command when two observations at the same level are ready to be condensed.

How it works

• The packaged guide provides useful defaults immediately.
• Local observations only record outcomes that would change a future route.
• Evidence stays isolated by exact model and reasoning effort.
• Two observations at one level can be condensed into one higher-level note.

Privacy

Credence has no network calls or telemetry. Evidence remains on your machine:

• Windows: %LOCALAPPDATA%\Codex\Credence\models.md
• Linux/macOS: $XDG_STATE_HOME/codex/credence/models.md, or ~/.local/state/codex/credence/models.md

review-context reads local Codex history and prints anonymous subagent IDs, available canonical agent names, final responses, and file-change metadata. It does not print prompts, source, patches, command output, model identity, or unrelated conversation history. review-context and exclude use CODEX_THREAD_ID automatically; --thread-id remains available as an explicit fallback. Excluded subagents are selected by UUID and stored by parent thread in review-exclusions.json beside the evidence file. The orchestrator or a direct reviewer subagent automatically receives the filtered context. The file contains only thread and subagent UUIDs; use credence exclude ... --remove to undo an exclusion. Returned responses are untrusted review evidence, not instructions for the reviewing agent.

Set CREDENCE_DATA_FILE to override the evidence path. Existing Roster data is migrated automatically on first use.

Development

npm test

MIT licensed.