name	inference-reconcile
description	Reconcile the inference-engine catalog. Re-fingerprints every artifact, re-runs Wald SPRT and Beta-Binomial and EMA decay, elevates patterns that cross the SPRT threshold, retires patterns that fall below it, and atomically rewrites catalog.json. Safe to run repeatedly — fully idempotent on identical artifact streams. Auto-triggers on: "/inference-reconcile", "rebuild the catalog", "re-elevate patterns", "refresh ufopedia", "reconcile learnings".
allowed-tools	Bash(python *) Read Agent

Inference Reconcile

Re-derive catalog.json from the full artifact history. Fully autonomous.

Usage

The caller provides no arguments at Phase 1. Future phases may accept scope filters.

Pipeline

Step 1: Spawn the reconciler agent

Delegate to the Sonnet-tier reconciler. The agent runs the engine, validates output, and re-renders briefings when verdicts change:

Agent(subagent_type="general-purpose", model="sonnet",
      prompt="Run the reconciler agent defined at
              wixie/plugins/inference-engine/agents/reconciler.md.")

Step 2: Parse the agent's report

The agent returns one line:

reconciled <N> artifacts -> <P> patterns (<E> elevated, <R> retired)

Step 3: Re-render briefings if verdicts changed

If the agent reports that verdicts changed (the agent diffs against the prior catalog internally), a fresh state/briefings/wixie.md is already written. Otherwise re-render unconditionally — cheap and keeps the briefing timestamp current:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/inference-engine.py render-briefing wixie

Step 4: Report to caller

Reconcile complete: <N> artifacts, <P> patterns (<E> elevated, <R> retired)
Briefing: state/briefings/wixie.md

If WIXIE_INFERENCE_ENABLED=0 the reconcile still runs (it's safe) but the emit pipeline is a no-op, so the catalog may not reflect recent sessions. Tell the caller honestly.

Rules

Do NOT edit catalog.json by hand at any step. The atomic write path in inference-engine.py reconcile is the only supported mutation.
Do NOT claim elevation for patterns that did not cross SPRT. Honest numbers are the product.
Do NOT skip the briefing refresh when verdicts change — stale briefings erode the substrate's value.
If the engine script is missing or errors, report the error verbatim and stop. Do not invent a result.

Inference Reconcile

Re-derive catalog.json from the full artifact history. Fully autonomous.

Usage

The caller provides no arguments at Phase 1. Future phases may accept scope filters.

Pipeline

Step 1: Spawn the reconciler agent

Delegate to the Sonnet-tier reconciler. The agent runs the engine, validates output, and re-renders briefings when verdicts change:

Agent(subagent_type="general-purpose", model="sonnet", prompt="Run the reconciler agent defined at wixie/plugins/inference-engine/agents/reconciler.md.")

Step 2: Parse the agent's report

The agent returns one line:

reconciled <N> artifacts -> <P> patterns (<E> elevated, <R> retired)

Step 3: Re-render briefings if verdicts changed

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/inference-engine.py render-briefing wixie

Step 4: Report to caller

Reconcile complete: <N> artifacts, <P> patterns (<E> elevated, <R> retired) Briefing: state/briefings/wixie.md

If WIXIE_INFERENCE_ENABLED=0 the reconcile still runs (it's safe) but the emit pipeline is a no-op, so the catalog may not reflect recent sessions. Tell the caller honestly.

Rules

Do NOT edit catalog.json by hand at any step. The atomic write path in inference-engine.py reconcile is the only supported mutation.

Do NOT claim elevation for patterns that did not cross SPRT. Honest numbers are the product.

Do NOT skip the briefing refresh when verdicts change — stale briefings erode the substrate's value.

If the engine script is missing or errors, report the error verbatim and stop. Do not invent a result.

inference-reconcile