gapmiss/claude-sessions/distill

/distill: live session distillation

Distill the current Claude Code session into a structured Obsidian note with queryable frontmatter.

CRITICAL: Output the markdown content directly in your text response. Do NOT use Write, Edit, or any file-writing tools. The user will copy the output themselves.

Workflow

1. Gather metadata: extract session_id, cwd, git branch, model, start time, token counts, cost, tool usage, and files touched from what is already visible in the conversation (system reminders, git status, prior tool results). Do NOT run Bash or other tools to fetch metadata. Use fallbacks for anything unavailable.
2. Classify session type: review the conversation and assign one or more types from the vocabulary below.
3. Synthesize content: extract key decisions (with reasoning), non-obvious learnings, and the most important exchanges from the full conversation.
4. Write output: output the markdown content directly in your response (NOT using Write or Edit tools). Include the suggested filename per the naming convention.
5. Validate output: verify the YAML frontmatter parses correctly and all required fields are present before finalizing. Confirm session_id, schema_version, project, cwd, and source_path are populated; check that session_type contains at least one valid vocabulary term.

Handling Missing Metadata

When a metadata field cannot be determined, apply the following fallbacks rather than omitting the field:

If start_time is unavailable, use the current timestamp and note it is approximate.

Output Format

Write a single markdown file to stdout with this exact structure:

---
tags:
  -claude-session
  -claude-session/{type}    # nested tag for each session_type
session_id: "{session_id}"
schema_version: 1
 
project: "{project_name}"
cwd: "{working_directory}"
branch: "{git_branch}"
model: "{model_name}"
title: "{custom_title}"          # only if session was renamed
 
start_time: {ISO8601_timestamp}
duration_min: {minutes}
 
cost_usd: {cost}
context_tokens: {context}
input_tokens: {input}
output_tokens: {output}
cache_read_tokens: {cache_read}
 
user_turns: {count}
assistant_turns: {count}
tools_used: [{tool1}, {tool2}, ...]
files_touched:
  -"[[relative/path/to/file.ts]]"
 
error_count: {number}
 
session_type:
  -{type1}
  -{type2}
 
source_path: "{absolute_path_to_jsonl}"
---
 
## Summary
 
2-3 sentences describing what was accomplished in this session.
 
## Decisions
 
-**Decision 1**: Why this choice was made
-**Decision 2**: Why this choice was made
 
## Learnings
 
-Non-obvious discovery or gotcha worth remembering
-Pattern that worked well or should be avoided
 
## Key Exchanges
 
> User: The important question or request...
> Assistant: The key insight or solution...

Session Type Vocabulary

Use one or more of these values for session_type:

bug-fix | feature | refactor | exploration | discussion | config | docs | test | review | deploy

Note Naming Convention

{project}--{YYYY-MM-DD}--{short-id}.md

• project: sanitized project name (non-alphanumeric → hyphens, lowercase)
• YYYY-MM-DD: from start_time
• short-id: first 8 characters of session_id

Example: claude-sessions--2026-04-08--a1b2c3d4.md

Frontmatter Field Rules

1. Exact values required for: session_id, schema_version, project, cwd, source_path
2. Approximate OK for: duration_min, cost_usd, token counts (Layer 0 will correct)
3. LLM-only fields: session_type (you classify), Summary/Decisions/Learnings/Key Exchanges sections
4. Wikilinks: files_touched uses [[path]] syntax for Obsidian linking