DeltaTorch

Save the task. Switch agent. Keep going when limits hit.

DeltaTorch is a small handoff skill + CLI for coding sessions that move between agents before context, usage, rate, token, or paid-plan limits stop the work. One agent saves a compact checkpoint. The next agent resumes from that checkpoint instead of redoing the same work or making you copy-paste context.

The goal is simple: save tokens, optimize how you use paid AI-agent plans, and fall back to local or free models when the expensive agent is close to its limit. It also works for planned agent roles: Claude can plan, Codex can implement, Claude can review.

Works with Claude Code, Codex, Copilot, and local agents too. The handoff is just Markdown in .handoff/.

npm

Install

Pick your agent. One command. Done.

| Agent | Install | | --- | --- | | Claude Code | npx skills add MethosPi/delta-torch -a claude-code -y | | Codex | npx skills add MethosPi/delta-torch -a codex -y | | GitHub Copilot | npx skills add MethosPi/delta-torch -a github-copilot -y | | Any other | npx skills add MethosPi/delta-torch -y |

Install once. After that, the agent should run DeltaTorch itself instead of asking you to copy commands.

Flow

Start a task with any agent. If that agent gets close to a token, usage, or rate limit, it saves the current task context so another agent can continue from the same point.

You can also split the work intentionally: one agent plans, another executes, and another reviews. The same checkpoint format works for both cases.

1. Start in Claude Code, Codex, Copilot, or a local Ollama-based agent.
2. The agent saves a handoff before it reaches a hard limit, or when you switch phase.
3. Open the next agent and say: continue task 3.
4. The next agent runs DeltaTorch, loads the checkpoint, and continues from the saved state.

Same project. Different agent. No restart. Fewer wasted tokens.

Quick Demo

$ pnpm dlx delta-torch init
Initialized handoff state in /path/to/project

$ pnpm dlx delta-torch save --reason context-limit <<'EOF'
## Goal
Finish the README positioning.

## Work Completed
Outlined the fallback messaging.

## Next Steps
Resume in Codex and edit the docs.
EOF
Saved task #1 (2026-04-26T12-30-00Z-context-limit)
Continue in another agent with: pnpm dlx delta-torch resume 1
/path/to/project/.handoff/handoffs/2026-04-26T12-30-00Z-context-limit.md

$ pnpm dlx delta-torch resume 1
Task: #1
Checkpoint: 2026-04-26T12-30-00Z-context-limit
File: .handoff/handoffs/2026-04-26T12-30-00Z-context-limit.md

# Agent Handoff

## Goal
Finish the README positioning.

## Next Steps
Resume in Codex and edit the docs.

## Resume Prompt
Continue this task from the saved Agent Handoff checkpoint.

Claude Plans -> Codex Implements -> Claude Reviews

# Claude Code saves the plan before handing off.
pnpm dlx delta-torch save --reason plan-complete

# Codex resumes the plan and implements.
pnpm dlx delta-torch resume 1

# Codex saves the implementation checkpoint for Claude review.
pnpm dlx delta-torch save --reason ready-for-review

The last command prints the next short task number, for example #2; Claude Code can resume that checkpoint for review.

What It Does

• Saves compact task checkpoints in .handoff/
• Gives each handoff a short task number like #3
• Lets another agent resume by task number, reason, or latest
• Saves before visible context, usage, rate, or token-budget limits leave work half-finished
• Helps preserve paid-plan quota by moving suitable work to cheaper or local agents
• Works for intentional planner/executor/reviewer handoffs, not only limit recovery
• Keeps the format simple and audit-friendly
• Avoids secrets from common private paths and token patterns

CLI

No global install required.

pnpm dlx delta-torch init
pnpm dlx delta-torch save --reason ready-for-review
pnpm dlx delta-torch resume 3
pnpm dlx delta-torch list

If you want the CLI in the repo:

pnpm add -D delta-torch

File Layout

.handoff/
  config.json
  registry.json
  handoffs/
    2026-04-26T12-30-00Z-context-limit.md

Each handoff stores:

• original prompt
• goal
• completed work
• current state
• changed files
• commands run
• risks
• next steps
• a ready-to-use resume prompt

.handoff/ Policy

Default recommendation: keep .handoff/ local and ignored. It is working-session metadata, not source code.

Commit .handoff/ only when your team intentionally wants shared handoff history in the repository. DeltaTorch does not edit .gitignore for you, but its own automatic git snapshot ignores .handoff/ metadata so checkpoint files do not recursively describe themselves.

Security Boundaries

DeltaTorch redacts common private paths and token patterns, including .env files, private key blocks, bearer tokens, URL credentials, and common secret assignment names.

That is a safety net, not a secret manager. Do not paste API keys, private keys, passwords, .env contents, customer data, or other secrets into checkpoint text.

Notes

• DeltaTorch is not an MCP server, not an orchestrator, and not a daemon.
• Claude Code gets slash-friendly skills. Other agents use the same handoff files and CLI.
• DeltaTorch does not read hidden provider quota telemetry. Agents save proactively from visible warnings, context meters, rate-limit messages, or token budgets you gave them.
• Releases are gated on v* tag pushes. Add a changeset per change with pnpm changeset, run pnpm changeset version to bump package.json and update CHANGELOG.md, then push the resulting tag to publish.

Manual Claude Install

If you want the Claude-only fallback:

pnpm dlx delta-torch install-skills --target project

License

MIT