Telemetry Terminology Similarity

Score pairwise similarity of telemetry field names across three independent layers. Emits raw scores — no thresholds, no clustering, no opinions. The consuming AI agent applies its own domain judgment.

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

When to Use This Skill

Use this skill when:

• Auditing a telemetry/logging schema for naming collisions
• Comparing two JSONL log schemas for field overlap
• Detecting

  trace_id

  vs

  traceId

  vs

  request_id

  vs

  correlation_id

  style problems
• Validating field naming consistency before shipping telemetry changes

Architecture

5-layer scoring pipeline — each layer catches what the others miss:

┌─────────────────────────────────────────────────────────┐
│  Layer 1: NORMALIZE                                     │
│  camelCase/snake_case split + abbreviation expansion    │
│  wordninja for concatenated words                       │
│  "traceId" → "trace id", "ts" → "timestamp"            │
├─────────────────────────────────────────────────────────┤
│  Layer 2: SYNTACTIC (RapidFuzz, 0-100)                  │
│  token_set_ratio on normalized forms                    │
│  Catches: trace_id ↔ traceId, level ↔ log_level        │
├─────────────────────────────────────────────────────────┤
│  Layer 3: TAXONOMIC (WordNet Wu-Palmer, 0.0-1.0)        │
│  Head-noun synonym detection via hypernym tree          │
│  Catches: level ↔ severity, error ↔ fault, op ↔ action │
├─────────────────────────────────────────────────────────┤
│  Layer 4: SEMANTIC (sentence-transformers, 0.0-1.0)     │
│  Cosine similarity via all-MiniLM-L6-v2 embeddings     │
│  Catches: error ↔ exception, user_id ↔ account_id      │
├─────────────────────────────────────────────────────────┤
│  Layer 5: CANONICAL (--canonical flag, optional)        │
│  RapidFuzz vs bundled OTel/OCSF/CloudEvents dictionary  │
│  Catches: http_method → http.request.method (OTel)      │
├─────────────────────────────────────────────────────────┤
│  Output: All pairs scored + canonical anchors.          │
│  Agent decides what to act on — tool computes, judges.  │
│  Use proposer-prompt.md for structured rename proposals.│
└─────────────────────────────────────────────────────────┘

Two-Phase Workflow: Score → Propose

The skill works in two phases:

1. Phase 1 — Score (

   term_similarity.py

   ): Compute raw similarity scores across 5 layers. Tool computes, no opinions emitted.
2. Phase 2 — Propose (

   references/proposer-prompt.md

   ): A bundled prompt template that consumes the scoring JSON and asks the LLM to produce structured rename proposals with confidence levels, evidence citations, and explicit escape hatches.

The two phases are deliberately separated. Phase 1 is deterministic and reproducible; Phase 2 applies domain judgment that only an LLM with conversation context can provide.

Dependencies

All installed via

uv run

sentence-transformers

rapidfuzz

wordninja

nltk

orjson

First run downloads the

all-MiniLM-L6-v2

Script Location

The analysis script lives in this skill's

references/

# SSoT-OK: marketplace path resolution for cross-repo invocation
SCRIPT_DIR="$(dirname "$(find ~/.claude/plugins -path '*/telemetry-terminology-similarity/references/term_similarity.py' -print -quit 2>/dev/null)")"
SCRIPT="$SCRIPT_DIR/term_similarity.py"

All examples below assume

$SCRIPT

Usage

Analyze field names directly

# SSoT-OK: uv run handles PEP 723 inline deps
uv run --python 3.13 "$SCRIPT" \
  trace_id traceId request_id correlation_id \
  level severity log_level priority

Extract fields from a Python codebase

Use Python regex extraction (macOS lacks

grep -P

python3 -c "
import re, glob
fields = set()
for f in glob.glob('**/*.py', recursive=True):
    text = open(f).read()
    for m in re.finditer(r'\"([a-z][a-z0-9_]*?)\":', text):
        fields.add(m.group(1))
for f in sorted(fields):
    print(f)
" | uv run --python 3.13 "$SCRIPT"

Analyze from stdin (pipe from jq, etc.)

head -1 telemetry.jsonl | jq -r 'keys[]' | uv run --python 3.13 "$SCRIPT"

Analyze a JSONL file's fields

uv run --python 3.13 "$SCRIPT" --jsonl /path/to/telemetry.jsonl

Compare two JSON schemas

uv run --python 3.13 "$SCRIPT" --schema-a schema_v1.json --schema-b schema_v2.json

Control output size

uv run --python 3.13 "$SCRIPT" --top 30 field1 field2 field3   # Top 30 pairs
uv run --python 3.13 "$SCRIPT" --top 0 field1 field2 field3    # All pairs
uv run --python 3.13 "$SCRIPT" --json field1 field2 field3     # JSON output

Lookup against canonical standards (OTel/OCSF/CloudEvents)

# Anchor each field against 1,453 bundled canonical names from OTel + OCSF + CloudEvents
uv run --python 3.13 "$SCRIPT" --canonical http_method http_status request_id severity

Output adds a

=== CANONICAL ANCHORS ===

Generate structured rename proposals (Phase 2)

After running with

--json --canonical

references/proposer-prompt.md

# Phase 1: Score
uv run --python 3.13 "$SCRIPT" --json --canonical [fields...] > analysis.json

# Phase 2: Apply proposer prompt (paste analysis.json into the template)
# The LLM produces structured proposals.json — review atomically

Parameters

--top

--canonical

--jsonl

--schema-a/-b

--json

Output Format

Text output (default)

Fields analyzed: 21
Unique after normalization: 19

=== EXACT DUPLICATES (after normalization) ===
  trace_id  ==  traceId
  timestamp  ==  ts

=== SCORED PAIRS (sorted by combined score) ===
    syn    tax    sem   comb  pair
    ---    ---    ---   ----  ----
  100.0  0.000  0.560  1.000  level                     <-> log_level
    0.0  1.000  0.472  1.000  error                     <-> fault
   66.7  0.909  0.457  0.909  operation                 <-> action
   46.2  0.833  0.251  0.833  level                     <-> severity
   28.6  0.667  0.700  0.700  error                     <-> exception

Three independent scores per pair — the agent reads all three to decide:

• syn (syntactic): high = surface-level name variant
• tax (taxonomic): high = WordNet synonym (hypernym tree)
• sem (semantic): high = embedding similarity (distributional)
• comb (combined): max(syn/100, tax, sem) — sorting key

JSON output (

--json

)

Structured JSON with

exact_duplicates

scored_pairs

How Each Layer Contributes

trace_id

traceId

level

severity

error

fault

operation

action

error

exception

user_id

account_id

Abbreviation Dictionary

The normalizer expands common telemetry abbreviations:

ts

uid

req

resp

err

msg

svc

env

op

lvl

evt

ctx

acct

cfg

dur

lat

Add domain-specific abbreviations by editing

ABBREVIATIONS

term_similarity.py

Troubleshooting

ModuleNotFoundError

uv run

$SCRIPT

grep: invalid option -- P

python3 -c "import re..."

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.