edercnj/x-telemetry-trend

Global Output Policy

• Language: English ONLY.
• Tone: Technical, Direct, and Concise.
• Efficiency: Remove all conversational fillers and greetings to save tokens.

Skill: Telemetry Trend

Purpose

Detect performance regressions across multiple epics using the append-only NDJSON telemetry logs. /x-telemetry-trend complements /x-telemetry-analyze: analyze answers "which phase is the bottleneck in epic X"; trend answers "is skill X getting slower over the last N epics?".

The skill consumes the global index at .claude/telemetry/index.json (on-demand; gitignored), aggregates P95 per skill per epic, and emits:

1. Top-10 regressions whose P95 delta vs. a moving baseline exceeds the configured threshold.
2. Top-10 slowest skills by mean-of-P95 across the analyzed window.
3. An interpretive observations block.

When to Use

• Periodic performance audit after a batch of merged stories.
• Before a refactor: confirm that a slow skill is trending up (not a fluke).
• In a CI job that should flag sustained regressions across epics.

Arguments

| Argument | Behaviour | | :--- | :--- | | --last N | Analyze the most-recent N epics (default 5). | | --threshold-pct P | Minimum delta % to classify as a regression (default 20). | | --baseline mean\|median | Baseline aggregation (default median — outlier-robust). | | --format md\|json | Output format (default md). | | --out path | Destination path. When omitted, writes to stdout (RULE-007). | | --base-dir path | Override plans/ (test-only). | | --index-path path | Override the cache path (test-only). | | --rebuild-index | Force rebuild ignoring the cache. |

Invocation

java -cp target/classes:target/dependency/* \
     dev.iadev.telemetry.trend.TelemetryTrendCli \
     --last 5 --threshold-pct 20 --baseline median

Exit Codes

| Code | Meaning | | :--- | :--- | | 0 | Success. Report emitted to stdout or --out path. | | 1 | Validation error (bad --last, --baseline, or --format value). | | 5 | Fewer than 2 epics with data — trend analysis needs a comparable series. | | 6 | --threshold-pct is negative. |

Algorithm

1. Build (or refresh) the per-skill per-epic P95 index by scanning plans/epic-*\/telemetry/events.ndjson. The index is persisted at .claude/telemetry/index.json and invalidated when an epic's NDJSON mtime changes.
2. Restrict to the most-recent --last N epics (natural ID order).
3. For each skill, compute the baseline (mean or median) of the N-1 oldest samples and compare against the current (latest) sample.
4. If deltaPct >= threshold, emit a regression entry; sort regressions by deltaPct descending, cap at top-10.
5. Rank skills by mean-of-P95 across the window; cap at top-10.

Output Layout (Markdown)

1. Header — generated-at, epics analyzed, threshold, baseline.
2. Top-10 regressions table (skill, baseline P95, current P95, delta%, epics).
3. Top-10 slowest skills table (skill, avg P95, invocations).
4. Observations — interpretive notes (regressions count, slowest skill).

Performance Contract

• 5 epics × 10 000 events → report in < 10 s (hard SLA, story §3.5).

The index builder uses TelemetryReader.streamSkippingInvalid() so memory stays bounded in the per-skill duration arrays.

Examples

# Default: last 5 epics, threshold 20 %, baseline median → stdout
/x-telemetry-trend

# Strict: threshold 10 %, mean baseline, write JSON to disk
/x-telemetry-trend --threshold-pct 10 --baseline mean --format json \
    --out plans/epic-0040/reports/trends.json

# Historical deep dive: last 10 epics, report to a custom path
/x-telemetry-trend --last 10 --out reports/quarterly-trends.md