Retrospective

Retrospective

Interactive post-session retro. Scans sessions, asks focused questions, proposes concrete actions the user approves in one step.

Modes

Single-session mode (default when inside a substantial conversation)

Scans the current conversation only. This is the original behavior.

Multi-session mode (default when invoked with no args, or with "today", or with a date)

Scans all sessions from a given day (default: today) across all projects. Extracts user corrections, skill failures, and patterns from JSONL transcripts.

Trigger: /retrospective at the start of a fresh session, or /retrospective today, or /retrospective 2026-05-24.

Multi-Session Discovery

Step 0 — Discover sessions

# Find today's sessions (default)
find ~/.claude/projects -maxdepth 2 -name "*.jsonl" -not -path "*/subagents/*" -mtime 0

# Or for a specific date, filter by file modification date
find ~/.claude/projects -maxdepth 2 -name "*.jsonl" -not -path "*/subagents/*" -newermt "YYYY-MM-DD" ! -newermt "YYYY-MM-DD + 1 day"

For each JSONL file found, extract a summary:

1. Parse the project name from the path (the directory name after projects/, decoded from the path-encoding)
2. Extract user and assistant messages (type user and assistant)
3. For user messages: message.content (may be string or array of {type: "text", text: "..."})
4. For assistant messages: collect text blocks from message.content array where type == "text"
5. Build a condensed transcript: first 10 user messages + last 5 user messages (to capture corrections at the end)
6. Skip sessions with fewer than 3 user messages (too short to have learnings)

Step 0b — Present session list

Show the user what was found:

Found N sessions today:
- project-name-1 (session-id[:8]) — "first user message preview..."
- project-name-2 (session-id[:8]) — "first user message preview..."

Then continue to Step 1a with the combined findings from all sessions. Each candidate action should note which session it came from (project name + short ID).

Single-Session Process

Step 0 — Gate Check (silent)

Scan the conversation and estimate session depth. Look for tool calls (Read, Edit, Write, Bash, Skill invocations), errors encountered, and back-and-forth exchanges. Don't try to count exactly — judge by feel:

• Short session (a quick question and answer, ~1-2 tasks) → Fast mode (Step 1b)
• Substantial session (multiple tasks, skill usage, errors, corrections) → Full mode (Step 1a)

Step 1a — Full Mode

Silently scan the conversation and collect:

1. Skills invoked — which succeeded, which failed, workarounds applied
2. User corrections — explicit "no, do it this way" moments (highest signal)
3. Repeated patterns — same error hit multiple times, same workaround applied
4. Cross-skill workflows — 3+ skills chained in sequence

Then read existing state:

• Find the current project's memory directory: glob ~/.claude/projects/*/memory/MEMORY.md and read it plus relevant memory files
• Read skill files for any skills that were invoked (~/.claude/skills/{name}/skill.md)
• Check if Linear CLI exists: test -f ~/.claude/skills/linear/scripts/linear && echo "configured" || echo "not configured"

Generate up to 5 candidate actions, ranked by signal strength:

1. User corrections (highest priority)
2. Failed/workarounded skills
3. Repeated patterns
4. Error patterns
5. Workflow patterns (lowest)

Dedup rules:

• If a candidate's content overlaps with an existing memory file → drop it
• If a skill update candidate overlaps with existing skill file content → drop it
• If Linear is not configured → omit any Linear task candidates

Present everything in a single AskUserQuestion call (up to 4 questions):

If Q1 = "Skip retro" → exit immediately.

If Q1 = "Rough" and Q2/Q3 are empty → exit with "Nothing to save — session closed." Don't add another question after the user already signaled they're done.

Step 1b — Fast Mode

Single AskUserQuestion call with one question:

• "Anything worth remembering from this session?" with options:
  • "Nothing, we're done" (default)
  • Other (free text)

If "Nothing" → exit. If free text → save as memory, exit.

Step 2 — Execute (silent, no re-confirmation)

For each approved item from Q4 (plus any insights from Q2/Q3 free text):

1. Read the target file before writing
2. Check for conflicts/duplicates against current content
3. Write the change if clean
4. Skip with warning if conflict detected

Action types and their targets: