oliver-kriska/session-deep-dive

Session Deep Dive (Tier 2)

Qualitative analysis of high-signal sessions identified by /session-scan. Spawns subagents with pre-computed metrics context for focused analysis.

Requirements

Requires ccrider MCP. If not available:

ccrider MCP is required. See: https://github.com/neilberkman/ccrider

Usage

/session-deep-dive ffa155ee-ed8a-492c-8797-878fcbec4d9e
/session-deep-dive --last                    # Most recent Tier 2 eligible
/session-deep-dive --from-scan               # All Tier 2 eligible from last scan
/session-deep-dive --from-scan --compare .claude/UPDATED_PLUGIN_REPORT_160_SESSIONS.md

Pipeline

Step 1: Resolve Target Sessions

From $ARGUMENTS:

• Session ID: Single session to analyze
• --last: Most recent Tier 2 eligible session from metrics.jsonl
• --from-scan: All sessions where tier2_eligible: true AND tier2_completed: false in .claude/session-metrics/metrics.jsonl
• --compare REPORT.md: Previous report to compare against (default: most recent .claude/session-analysis/insights-*.md)

If no metrics.jsonl exists, tell the user:

No metrics found. Run /session-scan first to discover and score sessions.

Step 2: Load Pre-computed Metrics

For each target session, read its entry from metrics.jsonl. Format the metrics as a context block for subagent prompts:

## Pre-computed Metrics (from /session-scan)

- Friction: 0.42 (retry_loops: 1, user_corrections: 3, approach_changes: 2)
- Fingerprint: bug-fix (confidence: 0.85)
- Plugin opportunity: 0.65 (could use: investigate, quick)
- Tool profile: Read 28.7%, Edit 15.2%, Bash 19.3%, Tidewave 22.8%
- Duration: 78 minutes, 19 user messages, 171 tool calls

Determine PROJECT_ROOT from current working directory.

Step 3: Fetch Transcripts — One Subagent Per Session

CRITICAL: One ccrider call = one subagent. Full transcripts are 5-30KB each. Even 3 per worker floods the worker's context.

For EACH session, spawn a haiku subagent:

Task(subagent_type="general-purpose", model="haiku", mode="bypassPermissions", prompt="""
Fetch one session transcript and save it.

1. mcp__ccrider__get_session_messages(session_id: "{SESSION_ID}")
   If > 200 messages: use last_n: 200

2. Write transcript to {PROJECT_ROOT}/.claude/session-analysis/{SHORT_ID}-transcript.md
   Format:
   # Session: {SHORT_ID}
   Project: {PROJECT}
   Date: {DATE}
   Messages: {COUNT}

   ## Messages
   ### User (seq N)
   {content}
   ### Assistant (seq N)
   {content}

3. Report: "Wrote {SHORT_ID}-transcript.md ({N} messages)"
""")

Spawn ALL fetch subagents in parallel. Wait for all to complete.

Step 4: Analyze Sessions

Read the analysis template — inline it into subagent prompts:

Glob: **/session-deep-dive/references/analysis-template-v2.md

ALWAYS use subagents — never analyze in main context.

• 1-6 sessions: Spawn sonnet subagents (one per session)
• 7+ sessions: Spawn haiku subagents for speed

Each analysis subagent prompt:

Read the session transcript at {transcript_path}. Apply the analysis template below to analyze this session. The pre-computed metrics below give you quantitative context — validate them and add qualitative depth.

{metrics_context_block}

{analysis_template_content}

Write your report (under 200 lines) to {report_path}.

Reports go to .claude/session-analysis/{short_id}-report.md.

Step 5: Compress (if 3+ sessions)

If 3+ sessions analyzed, spawn context-supervisor (haiku) to compress:

Read all report files in .claude/session-analysis/*-report.md. Write a consolidated summary to .claude/session-analysis/summaries/consolidated.md. Preserve: friction patterns, plugin opportunities, evidence strength tags. Remove: per-file details, generic observations, repeated context.

Step 6: Synthesize

Read the synthesis template:

Glob: **/session-deep-dive/references/synthesis-template.md

Read the --compare report (or latest insights file). Read MEMORY.md for known findings.

If 3+ sessions: read summaries/consolidated.md (NOT individual reports). If 1-2 sessions: read individual reports directly.

Produce synthesis comparing:

• New findings vs known patterns from MEMORY.md
• Confirmed patterns (seen before, still present)
• New patterns (not in previous reports)
• Resolved patterns (previously noted, no new occurrences)

Step 7: Update Ledger

Use Python to safely update metrics.jsonl — never manually read/modify/rewrite in the LLM context:

python3 -c "
import json
ids = {SESSION_IDS_SET}  # e.g., {'ffa155ee-...', '90a74843-...'}
lines = open('{PROJECT_ROOT}/.claude/session-metrics/metrics.jsonl').readlines()
with open('{PROJECT_ROOT}/.claude/session-metrics/metrics.jsonl', 'w') as f:
    for line in lines:
        entry = json.loads(line)
        if entry.get('session_id') in ids:
            entry['tier2_completed'] = True
        f.write(json.dumps(entry) + '\n')
"

Step 8: Write Output

Write synthesis to .claude/session-analysis/insights-{date}.md

Present key findings directly in conversation. Tell user:

Full report: .claude/session-analysis/insights-{date}.md Per-session reports: .claude/session-analysis/{id}-report.md

Output Files

| File | Purpose | |------|---------| | .claude/session-analysis/{id}-transcript.md | Raw transcript | | .claude/session-analysis/{id}-report.md | Per-session analysis | | .claude/session-analysis/summaries/consolidated.md | Compressed reports | | .claude/session-analysis/insights-{date}.md | Cross-session synthesis |

Iron Laws

1. ONE ccrider call = ONE subagent — never batch multiple fetches
2. NEVER fetch or analyze in main context — always subagents
3. Absolute paths in subagent prompts — subagents don't inherit skill context
4. Python for jsonl updates — never manually rewrite in LLM context
5. ALWAYS pass pre-computed metrics to analysis subagents — don't re-derive
6. NEVER skip synthesis — cross-session patterns are the real value
7. TAG evidence strength — every finding must be STRONG/MODERATE/WEAK