Session Log

Session Log

Scans today's agent transcripts from Cursor and Claude, extracts a brief summary of what was worked on, and writes a tagged daily entry into the Second Brain wiki. Composes with second-brain-ingest for wiki integration.

Skill Composition

What this skill adds: daily aggregation of all agent sessions across all workspaces (my-cursor-claw, prototype-creator, automated-usability-testing, cursor-skills) into a single structured wiki page with importance scoring, people detection, carry-forward tracking, and auto-resolve.

Workflow

Step 1: Determine the date

Default to today. If the user says "yesterday's sessions" or specifies a date, use that instead. Store as TARGET_DATE in YYYY-MM-DD format.

Step 2: Find agent transcripts for the date

Transcript locations:

Cursor transcripts (all workspaces):

• ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts/
• ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts/
• ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts/
• ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts/

Always scan all workspace transcript directories — work in prototype-creator, automated-usability-testing, and cursor-skills is core internship work and must be captured even when running session-log from my-cursor-claw.

Claude Code transcripts (if available): ~/.claude/projects/ — check for any JSONL files modified on TARGET_DATE.

List directories modified on the target date across all workspaces:

for dir in \
  ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts \
  ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts \
  ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts \
  ~/.cursor/projects/<YOUR_WORKSPACE_PROJECT_ID>/agent-transcripts; do
  find "$dir" -maxdepth 1 -type d -newermt "{TARGET_DATE} 00:00" \
    ! -newermt "{TARGET_DATE} 23:59" 2>/dev/null
done | sort

Also check for Claude Code conversations if the directory exists:

find ~/.claude/projects/ -name "*.jsonl" -newermt "{TARGET_DATE} 00:00" \
  ! -newermt "{TARGET_DATE} 23:59" 2>/dev/null | sort

Step 3: Read and summarize each session

For each transcript found, read the JSONL file and extract:

1. Topic — what was the user trying to do? Read the first user message (look for <user_query> tags or the raw message text).

2. What was worked on — scan assistant messages for tool calls that reveal actions taken:

   • Write / StrReplace tool calls = files created or modified
   • Delete tool calls = files deleted
   • Shell tool calls with git, gh, npm = commands run
   • CallMcpTool calls = external system interactions

3. What changed — extract file paths from tool calls to build a list of files created, modified, and deleted.

4. Why — read the assistant's reasoning around each change. Look for explanatory text between tool calls that describes intent.

5. Outcome — did the session end successfully? Check the last few messages for confirmation, errors, or open questions.

Reading strategy to keep it fast:

• Read the first 10 lines for the topic and initial context
• Scan for tool call patterns (Write, StrReplace, Delete, Shell, CallMcpTool) using ripgrep rather than reading every line:

  rg '"name":"(Write|StrReplace|Delete|Shell|CallMcpTool)"' <transcript.jsonl> | head -30
  

• Read the last 10 lines for the outcome
• Skip the middle unless the topic is unclear from the bookends

Cap at 15 transcripts per day.

Step 3.5: Score importance

After summarizing each session, assign an importance score (1–5) using these signals additively (start at 1, cap at 5):

Scoring rules:

• Minimum score is 1, maximum is 5.
• Sessions scoring 4–5 are marked [IMPORTANT] in the output.
• Carried-forward items that reach 3+ days are auto-promoted to score 5 regardless of other signals.

Step 3.6: Detect people

Scan each session's transcript content for references to key collaborators. Apply people tags when detected:

People tags use the pf-m-blue color class (matching tag-scanner conventions).

Step 4: Build the daily entry

Sort sessions by importance score (highest first). Create a structured summary with this format:

# Session Log — YYYY-MM-DD

**Signal:** <N> important items<, M stale carry-forwards if any>

**Tags:** #daily-log #YYYY-MM-DD #<people-tags> #<status-tags> #<project-tags>

## Summary

<2-3 sentence overview of the day's agent work, leading with the most important item>

## Sessions (sorted by importance)

### 1. [IMPORTANT] <High-score session topic>
- **Importance:** N/5 (<brief reason — e.g. "involves manager + deliverable">)
- **What:** <brief description of work done>
- **Changed:** <files created/modified/deleted>
- **Why:** <reasoning or motivation>
- **Outcome:** <completed / in progress / blocked>

### 2. <Medium-score session topic>
- **Importance:** N/5 (<brief reason>)
- **What:** ...
- **Changed:** ...
- **Why:** ...
- **Outcome:** ...

### 3. <Low-score session topic>
- **Importance:** N/5 (<brief reason — e.g. "networking, no deliverables">)
- **What:** ...
- **Outcome:** ...

## Carried Forward

### [DAY N] <Topic from previous day>
- **Originally:** YYYY-MM-DD
- **Importance:** N/5 (<auto-escalated if 3+ days>)
- **Last status:** <in progress / blocked — details>

## Files Touched

| File | Action | Session |
|------|--------|---------|
| .cursor/skills/work-context/SKILL.md | Created | 1 |
| canvases/work-context.canvas.tsx | Created | 1 |
| .cursor/skills/meeting-prep/SKILL.md | Created | 2 |

## Related

- [<linked wiki pages>](<file>.md)

Output rules:

• Sessions scoring 4–5 get the [IMPORTANT] prefix in their heading.
• Sessions scoring 1 can omit Changed and Why fields (just What + Outcome).
• The ## Carried Forward section only appears if there are carry-forward items.
• The **Signal:** line always appears — it's the at-a-glance summary for skimming.

Tag generation rules:

Always include:

• #daily-log
• #YYYY-MM-DD (the date)

People tags (auto-detected from Step 3.6, pf-m-blue color):

• #zack — any interaction, message, or reference to /
• #andy — references
• #beau — references
• #steven — references
• #team — group meetings, standups, retros (3+ people)

Status tags (derived from importance scoring and outcomes):

• #blocked (pf-m-orange) — session ended with unresolved blocker
• #shipped (pf-m-green) — deliverable completed and pushed/shared
• #carried-forward (pf-m-orange) — item appeared yesterday and is still open
• #stale (pf-m-orange) — carried forward 3+ days, auto-escalated
• #decision-made (pf-m-green) — a decision was recorded or committed to
• #manager-sync (pf-m-grey) — 1:1 or direct feedback loop with manager

Context tags (derived from content):

• #sprint-work (pf-m-purple) — touches an active Jira sprint ticket
• #networking (pf-m-grey) — intro meetings, coffee chats, no deliverables
• #skills — if any skill SKILL.md files were created or modified
• #canvas — if any .canvas.tsx files were created
• #eval — if eval-related work was done
• #mcp — if MCP tools were called for external system interaction
• #prototype-creator — if prototype-creator related work appeared
• #second-brain — if wiki pages were created or modified
• #jira — if Jira tickets were referenced or modified
• #github — if PRs or commits were made
• #debugging — if the session involved troubleshooting or fixing errors
• Add any project-specific tags that match existing wiki page slugs

Tag limits: 5–10 tags per entry. People and status tags do not count toward the limit — they are always included when detected. Prefer specific tags over generic ones.

Step 4.5: Carry forward unresolved items

Before writing, check for items that should carry forward from previous days.

Owner filtering (critical): Only carry forward items that belong to Evan or block Evan's work. Items tagged with [[The group]], [[Ansible UX]], [[EX team]], or any non-Evan owner are NEVER carried forward. They stay in the meeting note where they originated. Items tagged with another person (e.g. [[<COLLEAGUE_1>]]) are only carried if they explicitly block Evan's work.

1. Read yesterday's log: Open ~/second-brain/wiki/session-log-{YESTERDAY}.md (calculate YESTERDAY from TARGET_DATE). If it doesn't exist, check the most recent session-log file instead.

2. Extract unresolved items: Find sessions with Outcome: in progress or Outcome: blocked. Also look for an existing ## Carried Forward section in yesterday's log (items may already be carrying). Apply owner filtering.

3. Increment carry counter: For each unresolved item, track how many consecutive days it has appeared:

   • If the item's [DAY N] marker exists, increment N.
   • If it's new from yesterday (no marker), start at [DAY 2].

4. Auto-escalate stale items: If an item reaches DAY 3+:

   • Set its importance to 5 (regardless of other signals).
   • Add #stale tag to the daily entry.
   • Move it to the top of the Carried Forward section.

5. Auto-close dropped items: If a carried-forward item does NOT appear in today's transcripts AND is NOT tagged #blocked AND has been absent for 2 consecutive days, mark it Resolved (auto-closed — no activity for 2 days) and stop carrying it forward.

6. Auto-expire long-lived items: If an item reaches DAY 7+ with no progress noted in any transcript, mark it Expired (7+ days, needs re-evaluation) and stop carrying it forward. Move it to an ## Expired section for reference.

7. Build the Carried Forward section using this format:

   ## Carried Forward
   
   ### [DAY N] <Original session topic>
   - **Originally:** YYYY-MM-DD
   - **Importance:** N/5 (<reason — auto-escalated if 3+ days>)
   - **Last status:** <in progress / blocked — details from last log>
   

8. Skip if empty: If no items carry forward, omit the section entirely.

Step 5: Write to Second Brain wiki

Delegate to second-brain-ingest Steps 2-10:

1. Check existing pages — scan ~/second-brain/wiki/ first 10 lines of each .md file for cross-linking decisions (Step 2).

2. Tags — this skill already generates its own tags in Step 4 above (daily-log, people, status, context tags), so second-brain-ingest Step 3 can reuse them directly instead of re-deriving tags from scratch.

3. Create the wiki page at: ~/second-brain/wiki/session-log-YYYY-MM-DD.md

   Follow second-brain-ingest Step 4 rules:

   • Start with # Session Log — YYYY-MM-DD
   • Second line: **Tags:** #tag1 #tag2 ... from this skill's Step 4
   • Add a "Related" section with links to connected wiki pages
   • Add a "Sources" section noting the transcript UUIDs

4. Update existing pages — add a link back from any related pages in their "Related" section (Step 5).

5. Update index.md — add under a "Session Logs" category. Create the category if it doesn't exist yet (Step 6).

6. Update log.md — append a log entry (Step 7):

   ## YYYY-MM-DD — Session Log
   
   **Sessions:** N agent conversations
   **Created:** session-log-YYYY-MM-DD.md — <one-line summary>
   **Tags:** #daily-log #skills #canvas ...
   **Cross-linked with:** <list of updated pages>
   **Category:** Session Logs
   

7. Update the tag index — apply second-brain-ingest Step 8 so this entry's tags are queryable without re-scanning every wiki page.

8. Update the KB files — apply second-brain-ingest Step 9, plus two things specific to session-log:

   • Any item in this entry's ## Carried Forward section that's owned by Evan belongs in ~/second-brain/pending-actions.md too, not just the dated session log page -- carried-forward items are exactly the kind of thing that gets lost between a daily log entry and the next time anyone looks for it.
   • Overwrite ~/second-brain/rolling-context.md's ## Right Now section with today's highest-importance session topic(s) (the [IMPORTANT]-flagged ones), and refresh ## Recent Decisions if this session recorded any (#decision-made tag). This is the "daily journal writes back to the KB" step -- see second-brain-checkup for the complementary read side.

Step 6: Confirm

Tell the user:

• How many sessions were logged
• The wiki page path
• Which tags were applied
• Which existing wiki pages were cross-linked
• Any sessions that seemed incomplete or had open questions

Auto-Resolve Rules

This skill operates without user interaction by default. No "did you complete this?" prompts. The importance engine and carry-forward mechanism replace manual check-ins entirely.

What "auto" means in practice:

• Intro meetings with peers (no action items, no follow-ups) automatically score 1/5 and sink to the bottom. No tagging needed.
• A 1:1 with that produces a new task automatically scores 4/5 and gets [IMPORTANT] + #zack + #manager-sync.
• A blocked PR that carries for 3 days auto-escalates to 5/5 + #stale without any user action.

Fallback Behavior

Common Mistakes