Menú
Generates pipeline analytics report -- success rate, per-agent effectiveness, failure patterns, optional HTML dashboard
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Analyzes a specific bug from the issue tracker (analysis only, no code changes)
Validate Automation Config, MCP servers, and tokens
Creates backlog epics in issue tracker from a specification document
Multi-agent discussion -- brings 2-3 agent perspectives into one conversation
Run the fix pipeline on a single ticket OR a batch -- auto-resumes from checkpoint if state exists
Implements a feature from the issue tracker -- spec, design, fix, review, test, publish
| name | metrics |
| description | Generates pipeline analytics report -- success rate, per-agent effectiveness, failure patterns, optional HTML dashboard |
| allowed-tools | mcp__*, Read, Glob, Grep, Bash, Write |
| argument-hint | [--period <N>] [--output <path>] [--format <md|json|html>] |
Input: $ARGUMENTS = optional flags (--period <N>, --output <path>, --format <md|json|html>)
FORMAT="" # empty = no flag supplied = trigger interactive prompt path
PERIOD=30
OUTPUT=""
while [[ $# -gt 0 ]]; do
case "$1" in
--period)
PERIOD="$2"; shift 2
;;
--output)
OUTPUT="$2"; shift 2
;;
--format)
FORMAT="$2"; shift 2
[[ "$FORMAT" =~ ^(md|json|html)$ ]] || { echo "Error: --format must be 'md', 'json', or 'html'"; exit 1; }
;;
*)
shift
;;
esac
done
# Sentinel for interactive-prompt path
NO_FORMAT_FLAG=0
if [ -z "$FORMAT" ]; then
NO_FORMAT_FLAG=1
FORMAT="md" # default render is markdown; prompt fires AFTER render
fi
--period N → period in days (default: 30)--output path → output file (default: stdout)--format md|json|html → output format (default: md with interactive prompt)Read Automation Config from CLAUDE.md section ## Automation Config:
Before any pipeline operation, verify MCP tool availability:
mcp__* tool matching the tracker type is accessible/agent-flow:check-setup for diagnostics."Via MCP server (per Issue Tracker → Type), fetch all issues matching Bug query + Feature query (if it exists). For each issue: ID, title, state, comments, created date.
For each issue, go through comments. Extract:
Triage checkpoint:
Regex: ^\[agent-flow\] Triage completed\. Severity: (.+)\. Area: (.+)\.$
Spec checkpoint:
Regex: ^\[agent-flow\] Spec analysis completed\. Area: (.+)\. Criteria: (.+)\.$
Block comment:
Regex (multiline): ^\[agent-flow\] 🔴 Pipeline Block
Following lines: Agent: (.+), Step: (.+), Reason: (.+), Detail: (.+), Recommendation: (.+)
PR link:
Regex: PR: (https?://\S+) or #(\d+)
git log --oneline --since="{period} days ago" via Bash. Parse commit messages for issue ID references.
total_attempted = issues with a [agent-flow] comment in the given periodtotal_fixed = issues with PR merged (or in a state from State transitions → For Review/Done)total_blocked = issues with a [agent-flow] 🔴 Pipeline Block comment and still in Blocked statesuccess_rate = total_fixed / total_attempted (percentage)avg_time_to_fix = average time from the first [agent-flow] comment to merged PR (if data available)block_by_stage = count of blocks per pipeline stage (triage, analysis, fix, review, test, publish)top_block_reasons = top 5 most frequent "Reason:" values from block commentsFor each agent (analyst, fixer, reviewer, test-engineer, browser-agent, publisher, scaffolder, spec-writer, spec-analyst, architect, acceptance-gate, deployment-verifier, backlog-creator, sprint-planner, priority-engine, spec-reviewer, rollback-agent): count blocks, success rate, most frequent failure reason.
avg_fixer_iterations = from reviewer comments (how many REQUEST_CHANGES before APPROVE)failure_pattern_detection = if an agent blocks > 30% of issues → flag pattern with top reasonPer-issue estimate (heuristic fallback): count stages × model tokens (sonnet ~30k, opus ~50k, haiku ~5k per invocation). Total estimate for the period.
For each issue identified in Step 4, glob .agent-flow/*/state.json to locate per-run state files. Classify each pipeline run:
pipeline.total_tokens is present in state.json → use pipeline.total_tokens directly; do NOT apply heuristic constants.pipeline.total_tokens is absent (legacy run or incomplete pipeline) → apply heuristic constants from Step 6.tokens_used but no top-level pipeline.total_tokens → classify the pipeline as ESTIMATED at the pipeline level, but note which stages had measured data in the detail section.For each pipeline run, collect:
measured_tokens = pipeline.total_tokens if present, else 0measured_stages = count of stages where {stage}.tokens_used exists and is > 0estimated_tokens = heuristic sum for stages missing tokens_usedtotal_stages = count of all stages present in state.jsonMaintain two global accumulators across all issues:
all_measured_issues = list of issue IDs classified as MEASUREDall_estimated_issues = list of issue IDs classified as ESTIMATED (including HYBRID)Output format depends on --format flag (default: md).
When --format json: emit a single JSON object matching the schema below. Do NOT emit any markdown text. Write to --output path if specified, otherwise stdout.
JSON schema:
{
"generated_at": "ISO-8601 timestamp",
"period_days": 30,
"project": "string (tracker project key, e.g. PROJ — NOT full project name)",
"pipeline_overview": {
"issues_attempted": 0,
"issues_fixed": 0,
"issues_blocked": 0,
"success_rate": 0.0,
"avg_time_to_fix_hours": 0.0
},
"token_cost": {
"measured_issues": ["PROJ-42"],
"estimated_issues": ["PROJ-37"],
"measured_tokens": 0,
"estimated_tokens": 0
},
"block_analysis": {
"by_stage": [
{"stage": "triage", "blocks": 0, "pct": 0.0}
],
"top_reasons": [
{"reason": "string (sanitized — block.detail content excluded per state/schema.md hard contract)", "count": 0}
]
},
"per_agent": [
{
"agent": "fixer",
"invocations": 0,
"blocks": 0,
"success_rate": 0.0,
"top_failure": "string"
}
],
"recommendations": ["string"]
}
HARD CONTRACT — block.detail exclusion: top_reasons[].reason uses block.reason only (the sanitized 2-sentence summary from the block comment). block.detail is NEVER serialized into JSON output. This cross-references the comprehensive INCLUDE/EXCLUDE channel table in state/schema.md §"Sensitive field exclusion contract".
When --format md (default): emit markdown report as follows.
## Pipeline Metrics Report — {project} ({period} days)
### Pipeline Overview
| Metric | Value |
|--------|-------|
| Issues attempted | {N} |
| Issues fixed | {N} ({rate}%) |
| Issues blocked | {N} |
| Avg time to fix | {N} hours |
### Token Cost — Per Pipeline Breakdown
For each pipeline run in the period, emit separate line items for measured and estimated
tokens. NEVER emit a single combined grand total when any issues fall back to heuristics.
**Example output for a run with both measured and estimated data:**
Pipeline PROJ-42 (2026-04-18):
Measured: 42,150 tokens (8 stages)
Estimated: 12,500 tokens (2 stages, heuristic)
Total: 54,650 tokens
Pipeline PROJ-37 (2026-04-17): [ESTIMATED — legacy state]
Estimated: 85,000 tokens (heuristic: 2×opus + 1×sonnet + 1×haiku)
Pipeline PROJ-55 (2026-04-19): [MEASURED]
Measured: 78,300 tokens (5 stages)
Total: 78,300 tokens
**Hybrid pipeline (some stages measured, some not):**
Pipeline PROJ-60 (2026-04-20): [ESTIMATED — partial measurement]
Measured stages: triage (12,500 tok), code_analysis (9,800 tok)
Estimated stages: fixer_reviewer, test (heuristic)
Estimated: ~80,000 tokens
Note: pipeline classified as ESTIMATED because pipeline.total_tokens is absent;
2 of 4 stages had measured data.
### Token Cost — Period Summary
| Category | Issues | Tokens |
|----------|--------|--------|
| Measured (state.json) | {X} | {sum} |
| Estimated (heuristic) | {Y} | ~{sum} |
> Measured and estimated totals are NOT summed into a single grand total.
### Block Analysis
| Stage | Blocks | % of total |
|-------|--------|------------|
| triage | ... | ... |
| analysis | ... | ... |
| fix | ... | ... |
| review | ... | ... |
| test | ... | ... |
| publish | ... | ... |
Top block reasons:
1. {reason} ({N} occurrences)
...
### Per-Agent Effectiveness
| Agent | Invocations | Blocks | Success Rate | Top Failure |
|-------|-------------|--------|-------------|-------------|
| fixer | ... | ... | ... | ... |
| ... | ... | ... | ... | ... |
### Recommendations
- {threshold-based recommendations}
---
Data source: measured={X} pipelines (state.json.pipeline.total_tokens present), estimated={Y} pipelines (heuristic fallback).
**Provenance:** {X} pipeline(s) used measured token data from state.json (pipeline.total_tokens).
{Y} pipeline(s) fell back to heuristic estimates (sonnet ~30k, opus ~50k, haiku ~5k per stage).
{If Y > 0}: Pipelines run before per-stage usage tracking was introduced lack per-stage usage fields and are reported as estimated.
Estimated pipelines: {comma-separated list of estimated issue IDs and run dates}.
Generated: {timestamp} | agent-flow v{version}
If --output specified: write to file, otherwise stdout.
When $FORMAT == "html", generate a self-contained HTML file with inline CSS and JavaScript.
HTML escape (XSS defense for user-controlled data paths):
html_escape() {
local s="$1"
s="${s//&/&}"
s="${s//</<}"
s="${s//>/>}"
s="${s//\"/"}"
s="${s//\'/'}"
printf '%s' "$s"
}
Order is critical: & MUST be substituted first to avoid double-escaping subsequent entity references.
All 5 user-controlled / tracker-sourced fields MUST be passed through html_escape BEFORE being interpolated into the HTML output:
$(html_escape "$issue_title")$(html_escape "$state_label")$(html_escape "$block_reason")$(html_escape "$block_recommendation")$(html_escape "$timeline_content")HTML structure (ported from the deleted dashboard skill):
— when pipeline.total_tokens is absent from state.json[agent-flow] comments by timestamp desc)CSS (inline <style> block):
JavaScript (inline <script> block, minimal):
When $FORMAT == "html":
$OUTPUT is empty: HTML_PATH="./metrics.html"$HTML_PATH via Write toolWhen $FORMAT == "md" or $FORMAT == "json" AND $OUTPUT is non-empty:
$OUTPUT instead of stdoutWhen $NO_FORMAT_FLAG == 1:
After the markdown report is displayed on stdout, present the following Czech prompt to the user as a question requiring a response:
Save output? [1] No [2] JSON → stdout [3] HTML → ./metrics.html
Handle the response:
1 → exit (no save)2 → re-render the report at --format json to stdout (re-execute Step 7's JSON branch)3 → re-render the report at --format html to ./metrics.html (execute Step 7a + Step 8)1 (no save)When $NO_FORMAT_FLAG == 0: SKIP this step entirely. The user explicitly chose a format; honor that choice silently.
Autopilot safety: Autopilot does NOT invoke /metrics. The interactive prompt is therefore safe from non-TTY automation contexts.