A policy, not a framework
One command writes the policy into CLAUDE.md, between permanent markers. It updates in place and never duplicates. It can also add three subagents: advisor, executor and verifier.
← Back to abhirupbanerjee.com
- Open source
- MIT
- v3.2.0
- Claude Code plugin
- Personal side project
Delivery discipline for AI coding agents.
A Claude Code plugin that installs a standing operating policy. One writer. A verification gate. Failure loops capped at two cycles. Risky changes go through plan, execute, verify and review, each with its own subagent and model tier. Install it once, and every later session works this way.
the sandwich · Claude CodeREADME.md
5layers: prompt, context, harness, loop, graph
3subagents, each on its own model tier
1writer. Reviewers never edit.
2repair cycles at most, then it stops and asks
01 / IN PLAIN WORDS
What it does
AI agents write code faster than anyone can review it, and they declare themselves done. Stackwich turns plan, build, test and review into standing policy for Claude Code, so the agent works like a disciplined engineer.
Nothing done until observed
The verifier runs the project's real tests, lint or build and returns GATE: PASS or GATE: FAIL. It never repairs. A done report states the command run and what it printed.
Cheap models do the routine work
Model tiering is the cost control. The executor runs on a cheap tier. The stronger tier only plans and reviews. Established facts are not re-derived, and research runs in forks so it stays out of the main context.
Tools
Built first for Claude Code, with Grok and Codex versions. Versions for more AI tools, each using that tool's own models, are coming.
How I use it
In my program work, every eligible task runs through Stackwich's advisor loop: plan, execute, verify, review. It has helped every Scrum Master on the program. It is not used in developers' code generation.
02 / ARCHITECTURE
What gets installed, and where
From the repo layout and README. Dashed boxes are optional. Tap any box to see what it does.
Tap a box
Each component shows its role and the file it comes from.
Installed by the pluginOptional or separate port
03 / THE ADVISOR LOOP
One risky change, start to finish
The highlight moves through each step. Pause it or choose a step to explore. Every rule here is in the installed policy block or the agent files.
gate fail: back to executor, max 2 cycles
04 / MODEL PICKING
Which path, which model
This runs the policy's routing rules in your browser. Pick a tool, describe the change, and see which roles run and on which models. Then fail the gate and watch the loop cap. Each version uses its own tool's models.
Tool
Does the change do any of these?
Size and shape
05 / BUILD STATUS
What ships in 3.2.0
Checked against the public repo: README, CHANGELOG, plugin manifest, agent files, evals and CI.
In the release
- Claude Code plugin marketplace install, then one command: /stackwich:stackwich.
- Policy block at policy-rev 5, behind a marker that never changes, so upgrades happen in place.
- Three subagents: advisor, executor and verifier, with explicit input contracts.
- Single writer, capped failure loops, two-strike rule, and ceilings on every loop.
- Optional measured layer: a local event log, a report, and a four-person approval gate.
- Grok and Codex versions, each with its own revision and its own tool's models.
- Router integration: optional Stackwich verification in Dynamic LLM Router v0.4.0, per worker.
- Evals for five install paths and two policy behaviours. CI validates every change.
Not published yet
- Grok port update: use advisor, executor and verifier as separate roles when the runtime can spawn them.
Not built yet
- Picking the right model as it goes, to cut token cost.
- Versions for more AI tools, each using that tool's own models.
- More customization.
Install, from the README
/plugin marketplace add ovhirup/stackwich
/plugin install stackwich@stackwich
/stackwich:stackwich
Before you install
Both commands are typed inside a Claude Code session. The skill asks for scope, user level or this project only, then writes the policy and offers the three agents. Upgrade and uninstall steps are in the README.
06 / WITH DYNAMIC LLM ROUTER
How they fit together
Shipped in Dynamic LLM Router v0.4.0. It is off by default and turned on per worker.
Picks the tool
The router hands each task to a coding agent's command-line tool, such as Codex, Grok Build or Cursor, based on live quota.
Runs inside that tool
For a worker with Stackwich turned on, the tool runs plan, execute, verify and review in its own session, on its own models, and writes a report. Workers that do not support it run as before.
Commit only when it checks out
The router commits only on a passing gate, an approved review, an exact match between planned and changed files, and its own tests passing. One correction on the same worker, then the job is blocked for review. Verified commits are labelled Stackwich: verified.
Still to do
In my own setup, Codex and Grok Build each use their own Stackwich version. A live end-to-end check with Grok Build is pending. There is no Cursor version yet.