Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Session Analyzer (Standalone Utility)

Type: Standalone Utility Category: 0XX Shared

Analyzes a session for errors, inefficiencies, and improvement opportunities. Produces actionable fixes for skills, tools, hooks, and communication style.

Scope: Single-session deep analysis (10 dimensions). Broader than protocol self-audit, narrower than /audit-sessions batch. For skill self-audit: references/meta_analysis_protocol.md §7. For multi-day patterns: /audit-sessions.

When to Use This Skill

• After completing a task — find what went wrong and how to improve
• Periodically — audit recent sessions for patterns
• After a skill run — analyze how well the skill instructions worked
• When debugging tool/hook issues — find root causes in session data

Input

$ARGUMENTS:

• (empty) — analyze current session (conversation context)
• recent — scan latest JSONL session per agent (Claude, Codex)
• {skill-name} — focus analysis on that skill's execution within the session

hex-line is an optional accelerator for large local session logs and repo files. If MCP is unavailable, continue with built-in Read/Grep/Glob/Bash; this skill must never block on MCP connectivity.

Dimensions

Phase 1: Collect Session Data

Determine source based on $ARGUMENTS:

Current session (empty args)

Scan conversation context directly. You have access to all tool call results, errors, and messages.

Recent sessions (recent arg)

Find latest session per agent:

echo "=== LATEST SESSIONS ==="

echo "## Claude"
CLAUDE_LATEST=$(stat -c '%Y %n' ~/.claude/projects/*/*.jsonl 2>/dev/null | sort -rn | head -1 | cut -d' ' -f2-)
[ -n "$CLAUDE_LATEST" ] && echo "  $CLAUDE_LATEST ($(wc -l < "$CLAUDE_LATEST") lines)" || echo "  No sessions found"

echo "## Claude Active Sessions"
for f in "$HOME/.claude/sessions"/*.json; do
  [ -f "$f" ] || continue
  PID=$(basename "$f" .json)
  kill -0 "$PID" 2>/dev/null && echo "  ACTIVE: $(cat "$f")" || echo "  STALE: PID=$PID"
done

echo "## Codex"
CODEX_LATEST=$(stat -c '%Y %n' ~/.codex/sessions/????/??/??/rollout-*.jsonl 2>/dev/null | sort -rn | head -1 | cut -d' ' -f2-)
[ -n "$CODEX_LATEST" ] && echo "  $CODEX_LATEST ($(wc -l < "$CODEX_LATEST") lines)" || echo "  No sessions found"

Extract raw data from JSONL session

Run for each found session file ($F):

echo "=== TOOL CALLS ==="
grep -oE '"name"\s*:\s*"[^"]*"' "$F" 2>/dev/null | sed 's/"name"\s*:\s*"//;s/"//' | sort | uniq -c | sort -rn | head -30

echo "=== ERRORS ==="
grep -oE 'NOOP_EDIT|TEXT_NOT_FOUND|FILE_NOT_FOUND|HASH_HINT|DANGEROUS|out of range|mismatch|tool_use_error|permission denied' "$F" 2>/dev/null | sort | uniq -c | sort -rn

echo "=== TOOL LOOPS (3+ consecutive) ==="
grep -oE '"name"\s*:\s*"[^"]*"' "$F" 2>/dev/null | sed 's/"name"\s*:\s*"//;s/"//' | uniq -c | sort -rn | awk '$1 >= 3'

echo "=== BUILT-IN vs MCP ==="
grep -oE '"name"\s*:\s*"(Read|Edit|Write|Grep|mcp__hex-line__\w+)"' "$F" 2>/dev/null | sed 's/.*"name"\s*:\s*"//;s/"//' | sort | uniq -c | sort -rn

echo "=== HOOK EVENTS ==="
grep -oE 'hook_progress|blocking error|PreToolUse|PostToolUse|hex-confirmed|Obligatory use|Use mcp__hex-line' "$F" 2>/dev/null | sort | uniq -c | sort -rn

echo "=== SKILL/COMMAND INVOCATIONS ==="
grep -oE '"display":\s*"/[a-z-]+' "$F" 2>/dev/null | sed 's/"display":\s*"//' | sort | uniq -c | sort -rn

Treat keyword-based error counts as heuristic only. Confirm that reported matches come from actual tool failure/result events, not from prompts, templates, or referenced docs that contain the same strings.

Extract token statistics from Claude JSONL

Run for Claude session file ($F) to get real token usage:

echo "=== TOKEN STATS ==="
TOTAL_IN=$(grep -oE '"input_tokens":[0-9]+' "$F" | grep -oE '[0-9]+' | paste -sd+ | bc 2>/dev/null)
TOTAL_OUT=$(grep -oE '"output_tokens":[0-9]+' "$F" | grep -oE '[0-9]+' | paste -sd+ | bc 2>/dev/null)
CACHE_CREATE=$(grep -oE '"cache_creation_input_tokens":[0-9]+' "$F" | grep -oE '[0-9]+' | paste -sd+ | bc 2>/dev/null)
CACHE_READ=$(grep -oE '"cache_read_input_tokens":[0-9]+' "$F" | grep -oE '[0-9]+' | paste -sd+ | bc 2>/dev/null)
echo "in=${TOTAL_IN:-0} out=${TOTAL_OUT:-0} cache_create=${CACHE_CREATE:-0} cache_read=${CACHE_READ:-0}"

Phase 2: Analyze by Dimensions

Apply D1-D10 to collected data. Classify each finding:

D1: Tool Errors

Count by type: NOOP_EDIT, TEXT_NOT_FOUND, hash mismatch, out of range, FILE_NOT_FOUND. For each, identify which step/phase caused it.

D2: Tool Waste

• Reads without preceding outline (file >100 lines)
• Same file read 3+ times
• Bash cat/grep/head/sed when MCP equivalent exists

D3: Process Issues

• Tool loops: same tool on same file 3+ times without progress
• Retry storms: 3+ attempts with different parameters
• Dead ends: approach explored then abandoned
• Wrong target: looked at file X when needed Y

D4: Script Extraction

• Bash/Python scripts written during session
• If script would be needed on every run of the skill → candidate for references/scripts/

D5: Hook & Permission

• Hook blocks (PreToolUse/PostToolUse blocking errors)
• Built-in Read/Edit/Write/Grep used when MCP enforced
• hex-confirmed bypasses — were they justified?

D6: Communication

• Message length distribution (over-explaining vs terse)
• Unnecessary confirmations ("shall I proceed?")
• Missing status updates at milestones
• Redundant summaries

D7: Decision Quality

• Time to detect error and pivot (tool calls between error and correction)
• Trial-and-error sequences vs targeted approach
• Tool selection reasoning (used wrong tool first)

D8: Context Pressure

• Same file read multiple times across phases
• Information re-gathered that was already available
• Signs of thread loss (re-asking resolved questions)

D9: Subagent Quality

• Agent tool invocations with empty or low-value results
• Agent timeouts
• Overbroad prompts (agent did too much or too little)

D10: Scope Drift

• Initial goal vs actual deliverables — any deviation?
• Unplanned additions during execution
• Pivots without explicit reason

Phase 3: Map to Fixes

For each finding, determine specific target and fix: