Capture development environment state and upload to S3.

1. Generate manifest:

bash

uv run ~/.dotfiles/scripts/devenv-backup/devenv.py manifest > /tmp/devenv-manifest.json

Parse the JSON and read the uncommitted section.

1b. Display captured structure:

Show the user a tree view of all captured workspaces with their change IDs and bookmarks:

code

Captured structure:
├── pivot/
│   ├── default        @ knynrwpv
│   ├── compare-plots  @ xlksqkzq  [compare-plots]
│   └── tui-fixes      @ xnussuvl  [tui-fixes]
├── iam/
│   └── default        @ oklnutnl
└── dotfiles/
    └── default        @ ltxnrnzv

Format: workspace_name @ change_id [bookmark_name] (bookmark only if one exists)

If files were captured (manifest version 2), also show:

code

Files to backup:
├── ~/
│   └── notes.txt (1.2 KB)
└── ~/pivot/
    ├── architecture.md (5.4 KB)
    └── pro_critique/
        └── critique.txt (2.1 KB)

If symlinks were captured, show them:

code

Symlinks to backup:
├── ~/.jjconfig.toml -> .dotfiles/.jjconfig.toml
├── ~/.gitconfig -> .dotfiles/.gitconfig
└── ~/jj/CLAUDE.md -> .dotfiles/.claude/project-instructions/jj.md

Note on stale workspaces: If jj commands fail with "workspace is stale" errors, run jj workspace update-stale in that workspace directory first. The Python script handles this automatically, but you may encounter it when running manual jj commands.

2. Analyze uncommitted changes (if any):

If there are uncommitted changes, analyze them in parallel using subagents:

•For each uncommitted change, cd to the workspace directory, then run jj diff -r <change_id>
•Check if the change is empty (no file changes) AND has no description
•For non-empty changes: generate a suggested commit description based on the diff
•Default branch name = workspace name

Auto-skip empty changes: If a change has no file modifications AND no description (typically the current working copy @ commit), automatically skip it without prompting the user. These are just jj's empty working copy changes.

CRITICAL: Skip commits that ARE main: Before processing any change:

•Run jj log -r <change_id> in the workspace directory
•If the change's commit is directly on main (i.e., it IS main, not just branched from it), skip it entirely
•These commits are immutable and cannot be modified, described, or pushed to a new branch
•The manifest's uncommitted list may include the current working copy which happens to be sitting on main - these must be skipped

3. Present results to user:

Only show changes that need user action (non-empty AND (need a description OR have not been pushed to remote)):

code

| Workspace | Change ID | Current Description | Suggested Description | Branch | Action |
|-----------|-----------|--------------------|-----------------------|--------|--------|
| pivot     | xlksqkzq  | (empty)            | Add compare plots...  | compare-plots | push |
| iam       | efgh5678  | WIP                | Fix auth validation   | iam | skip |

If all uncommitted changes were auto-skipped (all empty with no description, or on main), skip directly to step 6.

Ask the user to confirm or edit:

•Description to use for commit
•Branch name for push
•Action: push (describe + push), discard (abandon), or skip (leave as-is)

4. Execute actions:

For each change based on user's decision, first cd to the correct workspace directory:

•push:

bash

cd <workspace_directory>

# First, check existing bookmarks
jj bookmark list

# Describe the change (if needed)
jj describe -r <change_id> -m "<description>"

# Handle bookmark:
# - If bookmark already exists and is "ahead by N commits": just push it
# - If bookmark exists and is synced: skip (already pushed)
# - If no bookmark exists: create it
# - NEVER set bookmarks on commits that are ancestors of main

# Only create if bookmark doesn't exist:
jj bookmark set <branch> -r <change_id>

jj git push --bookmark <branch>

•discard: jj abandon <change_id>
•skip: do nothing

5. Re-verify manifest:

Run manifest again:

bash

uv run ~/.dotfiles/scripts/devenv-backup/devenv.py manifest > /tmp/devenv-manifest.json

If uncommitted is still non-empty, show the remaining items and ask user if they want to proceed anyway or resolve them.

6. Get S3 base path from user:

Ask for:

•S3 base path (e.g., s3://bucket/users/sami@metr.org/)
•The backup will be stored at {base}/{machine}/{name}/
•Claude Code data will be stored at {base}/claude-code/{machine}/

Optional settings (mention defaults):

•--name: Backup name (default: today's date YYYY-MM-DD)
•--machine: Machine identifier (default: hostname)
•--agent-instructions: Optional freeform text to display during restore
•--timeout: Overall timeout in seconds (default: 120s)

7. Recommend dry-run first:

bash

uv run ~/.dotfiles/scripts/devenv-backup/devenv.py backup --base <s3-base-path> --dry-run

This shows what would be uploaded without actually uploading.

8. Run backup:

bash

uv run ~/.dotfiles/scripts/devenv-backup/devenv.py backup --base <s3-base-path>

# With optional flags:
uv run ~/.dotfiles/scripts/devenv-backup/devenv.py backup \
  --base <s3-base-path> \
  --name 2026-01-20 \
  --machine devpod \
  --agent-instructions "Run install.sh after restore"

# With custom timeout (default: 120 seconds):
uv run ~/.dotfiles/scripts/devenv-backup/devenv.py --timeout 300 backup --base <s3-base-path>

This single command:

•Generates the manifest (with files by default, root_dir from $HOME)
•Uploads manifest.json to {base}/{machine}/{name}/
•Uploads files to {base}/{machine}/{name}/files/
•Syncs Claude Code data to {base}/claude-code/{machine}/ (only session data, not git-tracked config or credentials)

9. Summary:

Print:

•Backup location
•Number of workspaces captured
•Number of files uploaded (if any)
•Number of symlinks captured (if any)
•Claude Code data synced
•Any uncommitted changes that were skipped

Done when: Backup is uploaded to S3.