Session Decisions

Session Decisions

Extract technical decisions and deliberations from Claude Code session history, organized for resume and portfolio use.

Two-stage output: Decision Log (default) → STAR Format (optional)

Arguments

(all optional)
--project <name>    Filter by project (partial path match)
--from <YYYY-MM-DD> Start date
--to <YYYY-MM-DD>   End date
--star              Additionally generate STAR format transformation
--all               Include ALL decisions (default: user-driven only)

Default: current project, all dates, user-driven decisions only

Execution Algorithm

Step 1: Scope Selection [MANDATORY]

digraph scope {
  "Arguments provided?" [shape=diamond];
  "--project given?" [shape=diamond];
  "Encode project path" [shape=box];
  "AskUserQuestion: scope" [shape=box];
  "Use current cwd" [shape=box];

  "Arguments provided?" -> "--project given?" [label="yes"];
  "Arguments provided?" -> "AskUserQuestion: scope" [label="no args, no cwd context"];
  "Arguments provided?" -> "Use current cwd" [label="no args, has cwd"];
  "--project given?" -> "Encode project path" [label="yes"];
  "--project given?" -> "Use current cwd" [label="no"];
}

Path Encoding: /Users/admin/Workspace/Socar → -Users-admin-Workspace-Socar

For partial project names, fuzzy match against ~/.claude/projects/:

ls ~/.claude/projects/ | grep -i "<project-name>"

Step 2: Find Session Files

PROJECT_DIR="$HOME/.claude/projects/<encoded-cwd>"
# Exclude subagent session files — only analyze main session conversations
SESSION_FILES=$(find "$PROJECT_DIR" -maxdepth 1 -name "*.jsonl" -type f 2>/dev/null)

# Date filtering (macOS) — use file modification time
for f in $SESSION_FILES; do
  MTIME=$(stat -f "%Sm" -t "%Y-%m-%d" "$f")
  # Keep only files within --from/--to range
done

Note: Use -maxdepth 1 to exclude subagents/ directory. Subagent sessions are spawned by main sessions and rarely contain independent decision-making.

Decision Tree:

Session files found?
├─ No → "No sessions found. Check project name or date range."
└─ Yes → How many?
    ├─ 1-5 files  → Skip to Phase 2 (deep analyze all)
    ├─ 6-50 files → Funnel Pipeline (Phase 1 scan → Phase 2 select)
    └─ 50+ files  → Warn + limit to recent 20 (confirm via AskUserQuestion)

Step 3: Phase 1 - Decision Signal Scan

Only runs for 6+ sessions. For 1-5, skip directly to Phase 2.

Extract each session's first message (display text) from history.jsonl, then score conversations for decision keywords.

# Extract session preview from history.jsonl
jq -r 'select(.sessionId == "SESSION_ID") | .display' ~/.claude/history.jsonl | head -1

# Extract conversation + keyword scoring per session
EXTRACT_SCRIPT="$HOME/.claude/plugins/cache/team-attention-plugins/session-wrap/1.0.0/skills/history-insight/scripts/extract-session.sh"

for session in $SESSION_FILES; do
  # Extract conversation (removes 94% metadata)
  EXTRACTED=$("$EXTRACT_SCRIPT" "$session" 2>/dev/null || echo '{"messages":[]}')

  # Get message count for normalization
  MSG_COUNT=$(echo "$EXTRACTED" | jq '.message_count // 1')

  # Keyword scoring — pipe through `strings` to handle control characters in JSONL
  RAW_SCORE=$(echo "$EXTRACTED" | strings | \
    grep -ciE 'architect|design|pattern|versus|vs|tradeoff|migration|instead|constraint|workaround|refactor|performance|optimize|compare|alternative|approach' || echo 0)

  # Normalize: score per 10 messages (avoids longer sessions dominating)
  NORM_SCORE=$((RAW_SCORE * 10 / MSG_COUNT))
done

Important: Real session JSONL files often contain control characters (U+0000-U+001F) that cause jq parse failures in the text-joining pipeline. Always pipe through strings before grep for the keyword scoring step.

Signal Classification (details: ${baseDir}/references/decision-heuristics.md):

• HIGH (normalized 5+): Target for deep analysis
• MEDIUM (normalized 2-4): Include if fewer than 10 HIGH sessions
• LOW/NONE (normalized 0-1): Skip

Step 4: Phase 2 - Deep Decision Extraction

Analyze full conversations from HIGH-signal sessions (up to 10).

Step 4a: User Attribution Analysis [CRITICAL]

Default behavior extracts ONLY user-driven decisions. Use --all to include Claude-proposed decisions.

For each session, separate role: "human" messages from role: "assistant" messages. Only human messages determine attribution.

Attribution Classification (details: ${baseDir}/references/decision-heuristics.md):

User-driven decision signals in role: "human" messages:

• Explicit choice: "A로 해", "이걸로", "2번", selecting from presented options
• Rejection: "그거 말고", "다시", "아니", "이건 아닌 것 같아"
• Direction: "처럼 해", "를 참고해", "패턴 써", "방식으로"
• Correction: "이게 아니라", "~로 바꿔", modifying Claude's output
• Specification: naming a specific pattern, library, convention, or reference

NOT user decisions (exclude by default):

• Simple task delegation: "~해줘" without specifying how
• Passive approval: "ㅇㅇ", "좋아", "진행해", "ㄱㄱ"
• Claude's own technical choices during implementation

Step 4b: Decision Point Extraction

Conversation patterns to detect:

1. Question-Exploration-Conclusion: Question asked → alternatives explored → conclusion reached
2. Rejection-Revision loop: First approach fails → alternative chosen
3. Pros/Cons enumeration: Explicit listing of advantages and disadvantages
4. "Instead"/"Rather" markers: Expressions indicating choice between alternatives

Extract from each decision point:

• Context: What problem was being solved?
• Options Considered: What alternatives were discussed?
• Decision: What was chosen?
• Rationale: Why was this chosen over alternatives?
• Outcome: What was the result? (if visible in session)
• Technical Details: Specific technologies, patterns, API names
• User Quote: Exact user message that drove the decision (original language)
• Attribution: user-directed / user-rejected / user-feedback / claude-proposed / claude-autonomous

Step 5: Phase 3 - Synthesis & Output

Group by category:

• Architecture Decisions (AD): Module structure, design patterns, layer separation
• Technology Selection (TS): Library/framework/API choice, migration strategies
• Problem-Solving (PS): Debugging, performance optimization, overcoming technical challenges
• Process Improvements (PI): Workflow, automation, development process changes

Generate output files:

mkdir -p ~/Workspace/work-hub/decisions

Decision Log → ~/Workspace/work-hub/decisions/YYYY-MM-DD-session-decisions.md

With --star flag, additionally generate: STAR Resume → ~/Workspace/work-hub/decisions/YYYY-MM-DD-star-resume.md

Templates: See ${baseDir}/references/output-templates.md

STAR transformation rules:

• Combine related decisions from same project/feature into one Story
• Emphasize "why" reasoning in the Action section
• Insert [TODO: add metrics] placeholder where quantitative results are not available
• Add Keywords section with technology terms for ATS optimization

Step 6: Report Results

## Session Decisions Extraction Complete

- **Sessions scanned**: {total} ({project}, {date_range})
- **Decision-rich sessions**: {high_count}
- **Decisions extracted**: {decision_count}
- **Categories**: AD({n}), TS({n}), PS({n}), PI({n})
- **Output**: ~/Workspace/work-hub/decisions/{filename}

Error Handling

Security Notes

• Never expose full filesystem paths in output (use ~ prefix)
• Be cautious of sensitive data in sessions (API keys, passwords)

Related Resources

• ${baseDir}/references/decision-heuristics.md — Decision detection patterns in detail
• ${baseDir}/references/output-templates.md — Decision Log + STAR templates
• session-wrap:history-insight — Session history access patterns