sindri-forge 0.7.0

Claude Code plugin core: bounded, target-driven optimization loops

sindri

A Claude Code plugin for bounded, target-driven optimization loops.

In the myth, Sindri is the dwarf smith who forged Mjölnir by iterating under adversity — each hammer-strike tested, kept only if it survived. This plugin is the same pattern: make a change, benchmark it, keep it if it's better, revert otherwise, repeat until the target is hit or the pool of ideas is exhausted.

Inspired by karpathy/autoresearch and davebcn87/pi-autoresearch, adapted to Claude Code.

One commit per kept experiment. Git is the audit log. The forge runs one way.

Install

/plugin install sindri@claude-community
/reload-plugins            # or relaunch Claude Code; commands are then live

Then, in any repo you want to optimize:

/sindri:forge reduce bundle_bytes by 15%

That's it. No pip install, no virtualenv. The Python backend (sindri-forge) is fetched on demand by uv. If uv isn't installed, /sindri:forge prints a one-line install command and stops.

Sindri is published in the Anthropic community marketplace (anthropics/claude-plugins-community). If you haven't added it yet: /plugin marketplace add anthropics/claude-plugins-community.

Requirements:

• uv — runs the backend in an isolated environment
• git ≥ 2.30
• gh CLI (authenticated) — needed only for sindri-finalize PR creation
• Claude Code with Task + ScheduleWakeup tools

First-run dogfood: walk through docs/DOGFOOD.md on a throwaway repo before pointing sindri at real code — it verifies every skill path and catches environment gaps before you commit an overnight run.

Quickstart

# In the repo you want to optimize:
/sindri:forge reduce bundle_bytes by 15%

Sindri will:

1. Scaffold .claude/scripts/sindri/benchmark.py if missing (interactive; asks you to describe the measurement in natural language).
2. Scan your repo and draft a pool of 10–30 candidates.
3. Ask you to approve / edit the pool.
4. Check out a new branch sindri/reduce-bundle-bytes-15pct and run the baseline 3×.
5. Loop autonomously — one experiment per wakeup, ~60s between them.
6. When the target hits (or the pool drains with ≥1 win), push the branch and open a PR.

Check in anytime with /sindri:status. Halt with /sindri:stop. Abandon with /sindri:clear.

Commands

Type /sindri: in Claude Code and every action is right there in the autocomplete menu — eight discrete, self-describing commands (no router, no guessing which sub-verb exists):

> /sindri:
  /sindri:forge     <goal statement>   Start a new optimization run from a goal — bounded,
                                        target-driven experiments on any deterministic metric.
  /sindri:status                       Print the status of the current run (or report none active).
  /sindri:stop                         Signal the current run to halt at its next wakeup.
  /sindri:clear                        Abandon the current run — destructive; deletes state + branch.
  /sindri:setup                        Scaffold or repair the benchmark script for a run.
  /sindri:help                         Show the command list and a one-line overview of each.
  /sindri:loop                         ADVANCED. Run one orchestrator tick manually, no reschedule.
  /sindri:finalize                     ADVANCED. Force a push + PR from the run's kept commits.

forge, status, stop, clear, setup, and help cover the normal lifecycle. loop and finalize are advanced/debug paths — the orchestrator runs them automatically during a live run.

Architecture

• Python core (src/sindri/) — pure functions: statistics, state file I/O, git wrappers, pool ordering, termination predicates, PR body rendering.
• Skills (skills/sindri-*/SKILL.md) — prose prompts for the decisions that need Claude's inference: scaffolding a benchmark, proposing a candidate pool, orchestrating the loop.
• Subagent prompt (prompts/experiment-subagent.md) — the contract every experiment subagent follows. Fresh context per experiment, no context bleed.

See docs/superpowers/specs/2026-04-19-sindri-design.md for the full design.

Development

uv pip install -e ".[dev]"
pytest                                            # all test layers
pytest -m "not e2e" -q                            # skip the slow end-to-end loop test
pytest tests/unit/test_plugin_artifacts.py -v     # just artifact validity checks
./scripts/smoke.sh                                # plumbing smoke in a throwaway repo

Local plugin install (contributors)

git clone https://github.com/4KMetrics/sindri.git ~/src/sindri
cd ~/src/sindri && ./scripts/install-plugin.sh   # symlinks the plugin only

To test local backend edits (instead of the published package), point the launcher at your checkout before running /sindri:forge:

export SINDRI_FORGE_SOURCE="$(pwd)"

License

MIT