weekly-reality-capture

name	weekly-reality-capture
description	Auto-generate a reality-first weekly log from daily memory logs, git history, and agent state. Uses Claude to synthesize what actually happened — not a strategic filter, not a curated narrative. Supports backfill by week number. Use when a user says "weekly recap", "what happened this week", "weekly capture", or "reality check".
metadata	{"emoji":"📸","category":"memory","platform":"any (requires Anthropic API key)"}

Weekly Reality Capture

Synthesize a reality-first weekly log from multiple data sources. The key insight: most weekly reviews optimize for narrative. This one optimizes for truth.

Data Sources

Source	What it captures
Daily memory logs	What the agent did/discussed each day
Git log	What code actually changed
Agent session state	Open threads, pending follow-ups
Open threads JSON	Carried-over work items

Usage

# Current week
python3 scripts/capture.py

# Specific week
python3 scripts/capture.py --week 2026-W14

# Dry run (print synthesis prompt, don't call Claude)
python3 scripts/capture.py --dry-run

Configuration (env vars)

Variable	Default	Purpose
ANTHROPIC_API_KEY	(required)	Claude API key for synthesis
CAPTURE_MODEL	claude-sonnet-4-5-20250514	Model for synthesis
MEMORY_DIR	~/agent-workspace/memory	Daily log directory (YYYY-MM-DD.md files)
WEEKLY_DIR	~/agent-workspace/weekly	Output directory for weekly captures
REPO_DIR	~/agent-workspace	Git repo to pull log from
SESSION_STATE_PATH	(optional)	Path to session_state.json
OPEN_THREADS_PATH	(optional)	Path to open_threads.json

Output Format

Each weekly capture produces YYYY-WNN.md:

# Week 14 — 2026-03-31 to 2026-04-06

## What Actually Happened
- ...

## Decisions Made
- ...

## Unresolved / Carried Over
- ...

## Patterns Noticed
- ...

Why "Reality-First"?

Most automated weekly summaries read like press releases — they highlight wins and bury problems. This tool is designed to report what actually happened, including:

• Threads that stalled without resolution
• Promises made but not delivered
• Work that was started then abandoned
• Time spent on ops vs. creative work

The synthesis prompt explicitly instructs Claude to resist narrative smoothing.

Cron Setup

Run Sunday night to capture the full week:

0 23 * * 0 /usr/bin/python3 /path/to/weekly-reality-capture/scripts/capture.py >> /tmp/weekly-capture.log 2>&1

Scripts

• scripts/capture.py — Main script. Collects data, builds synthesis prompt, calls Claude, writes output.