inference-emit

name	inference-emit
description	Append a single artifact (failure observation, correction, precedent) to the inference-engine's append-only stream. Use when a cross-session-relevant event occurs — a self-caught failure mode, a corrected misunderstanding, a precedent worth compounding. The emit is a no-op unless WIXIE_INFERENCE_ENABLED=1 is set. Auto-triggers on: "/inference-emit", "emit this to ufopedia", "log this as a precedent for future sessions", "record this failure pattern".
allowed-tools	Bash(python *) Read Write

Inference Emit

Append one artifact to wixie/plugins/inference-engine/state/artifacts.jsonl.

Usage

The caller provides either:

• A JSON record via stdin or a file path.
• Enough structured text to let you build the record.

Required fields

Field	Type	Example
code	string	F07, OP05, H01
category	string	process-discipline, branding-drift
title	string	short failure title, ≤ 120 chars
cause	string	one to three sentences
counter	string	the rule that prevents recurrence
signal	string	one sentence the reader applies next time
tags	string[]	lowercase, underscore or hyphen

Optional fields

Field	Type	Purpose
evidence	object	sub-session recurrence counts (see below)
scope	string	plugin or sub-plugin
source_session	string	human-readable session id

Evidence keys that boost SPRT

If the artifact documents multiple independent recurrences inside one session, set one of:

• evidence.iterations — build-loop iterations
• evidence.user_rounds_of_pushback — user corrections in a session
• evidence.occurrences — generic count
• evidence.times_hit — alias

Each N > 1 contributes N SPRT observations, not 1. Use the honest count.

Pipeline

Step 1: Construct the record

If the caller gave you a JSON record, use it. If they gave structured text, build the JSON yourself using the field spec above. Never fabricate evidence counts — ask if unclear.

Step 2: Emit

WIXIE_INFERENCE_ENABLED=1 python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/inference-engine.py emit <(cat <<'EOF'
<your JSON record>
EOF
)

Confirm the stdout line emitted <CODE> -> artifacts.jsonl.

Step 3: Optional reconcile

If the artifact is high-confidence (existing pattern with fresh evidence), suggest running /inference-reconcile. Do not auto-trigger reconcile from emit — the brand contract says hooks inform, they don't decide.

Step 4: Report

Tell the caller:

Emitted <code> to artifacts.jsonl
Fingerprint: <first 16 chars of SHA-1>
Next: /inference-reconcile when ready to update the catalog.

Rules

• Do NOT emit without WIXIE_INFERENCE_ENABLED=1. The engine's emit path short-circuits anyway; the skill reports the no-op honestly.
• Do NOT fabricate fields. If the caller's text is missing signal or counter, ask.
• Do NOT overwrite an existing artifact. The stream is append-only by design.