azurewraith

Three layers working together:

UserPromptSubmit hook: injects state context into every prompt... phase, available tools

Instructions: the model sees transition() and get_state() as MCP tools it can call. PreToolUse hook blocks disallowed tools at execution time and returns an error with what IS available.

PostToolUse hook: detects file changes and fires interrupts.

Things in CLAUDE.md are merely suggestions, hooks are enforcement. The model can rationalize away instructions but it can't rationalize away a blocked tool call.

The Pi plugin takes this further... pi's extension API lets us remove tools from the model's view entirely via setActiveTools() so it never even sees what it can't use. smaller tool space, better reasoning. Better Pi support is a recent enhancement and it's getting better.

Here's a week 2 update. a lot shipped since this post...

OSS: The `engine` and `agent` crates are now fully open (Apache 2.0). only the `gateway` and the various plugins are FSL with a 3-year clock. You can run the full state machine locally, self-hosted, no cloud dependency. The end-to-end workflow works out of the box with ollama and any 13B+ model now.

Multi-agent support: I validated and enhanced plugins for Codex CLI, Oh-My-Codex, and Pi alongside the existing Claude Code plugin. Same gateway, same workflows, different agent frontends. the Pi plugin in particular is interesting because pi's extension API supports things Claude Code doesn't yet... programmatic model switching and per-state tool filtering where the model literally never sees disallowed tools.

Interrupts: Reactive file triggers that force state transitions. model edits a migration file? interrupt fires, pulls it into a review state. it uses the History State pattern to return where it was in the state machine after.

Fork/join: Parallel sub-agent execution. planning state dispatches N implementation branches each in their own worktree, join collects results.

Allowed_commands: Per-state bash command restrictions. testing state can run pytest but not rm -rf. enforced in the hook, not the prompt.

Tangentally, the Forge post this week (https://news.ycombinator.com/item?id=48192383) validated the same thesis from a different angle... structural guardrails on small models outperform unconstrained frontier models. three independent projects converging on "the harness is first-class infrastructure" in roughly two weeks.

Next on the agenda: per-state model routing. use a local 12B for grunt work, route to Opus/GPT-5 for the one call that matters. the cost math is trending towards ~80% reduction on a 6-phase workflow

update: I released the of the crates (with the exception of the gateway as FSL) as Apache 2.0 today (and uploaded to crates.io) along with a UI piece that's Apache 2.0 as well.

You should be able to run self-hosted end-to-end with docker-compose, let me know with a GitHub issue if something isn't working in that regard

Interestingly enough we have found the same net result -- structural guardrails are the unlock for smaller models. Our approach in particular layers three things: a parse rescue for malformed/incorrect tool calls (similar to your retry nudges), content-level intervention (diff size rejection, checkpoint forcing) and state machine enforcement on top (per-phase tool restriction, transition guards). On 13B models we saw completion of a selection of SWE-bench tasks went from ~20% to 100%. With frontier models we saw a reduction in API calls from reduced thrashing.

One of the most surprising findings was when a 9B model self-corrected through 4 tool parse failures within the guard rails. It tried to use a complex tool (patch_file), kept failing and eventually downshifted to a simpler tool (edit_line) that it could actually execute. The guardrails didn't make the model smarter, it just narrowed the execution space until it could find something that worked.

Brief: https://statewright.ai/research

yeah... I wanted the smaller models to behave more like their state of the art siblings and was a little shocked that it just didn't work well... so I made small models work better. Having the larger models work better was just an added benefit

shipped ^_^

Hey it's me again. Some things that didn't fit in the README or the original post -- less about features, more about where this goes.

The plan/implement/test workflow is very basic and represents the most common agentic use case. But the state machine pattern applies to any multi-step work where agents are useful but susceptible to death spirals, hallucinations, or other non-deterministic quirkiness. This also enables Claude Desktop and other non-coding agents to perform useful constrained work.

I've been building a content pipeline for tabletop publishing and tested it a bit earlier yesterday. A research phase gathers lore and game details from a compendium, a drafting phase generates structured content including schema-specific JSON validation (so my Lua+LaTeX templates work without iterating). A review gate has me editing content directly (tmux+neovim dialog is great for this). The agent shapes the content, makes sure it conforms to JSON validation and content requirements, then I write it. Before I adapted the state machine to it, the agent tried to do everything all at once — calling multiple agents is sometimes effective but details get lost and you definitely lose visibility in the summarization. The state machine runs everyone serially (for now) but chaining and parallelization are on the roadmap.

While working with statewright on a different workflow over the weekend and Claude (as Claude does) attempted to write an intricate bash script to work around a guardrail... and statewright blocked it! I think that was when I knew there was some real power behind what's been built here. Enforcement has to be structural, not advisory.

Also, being generally useful for things besides coding you can start to think about things like SOC 2 change management. Every change needs a plan, a human review gate, audited implementation, pull request, review, human approval, and then finally a human to approve a production deployment. Today teams enforce this with checklists and hope. An agent constrained by a workflow that won't let it deploy without all the prerequisite pieces is enterprise delivery with an auditable paper trail and humans injected for approvals where they need to be - not managing each change's lifecycle.

The piece I'm most excited about is agent-generated workflows. You solve a problem once and maintain your context, then point the agent at the JSON schema and it creates and uploads a new workflow to statewright automatically that you can use immediately. No fine-tuning, no exhaustive prompt engineering, no dozens of agents... best-fit lightweight guardrails that agents help build themselves, compiling your intent into structure the models can't weasel their way out of. This is a fundamentally different reality than what the current state of the art is practicing. I think that's a big deal.

you might want to add an actual exclusion to the patents/licenses for those groups of people

Done :) https://github.com/statewright/statewright/blob/main/PATENTS...

Thank you for calling my attention to this. I think the world would be a better place if everyone just left the lawyers out of everything

I feel you on the Claude pulsing thing. Running (or trying to) run Opencode with any model I could throw at it to perform useful work like the frontier/proprietary models do (a tall order I know) is where I started. everyone makes the problem bigger (massive contexts, massive number of parameters, more transformers (MoE)) but I started with how can we make small/local LLMs perform better (do more with less)

Opencode's plan/build is decent, like it is in Claude Code... state machines are the next evolution. model agnostic, tooling agnostic (where feasible)

the Cargo.toml covers the built Rust crates (engine, agent). the plugins/ directory has it's own LICENSE.md with the FSL terms. split license: the engine is completely open source, plugins FSL with a 3 year clock. I should make this clearer in the workspace config. I am planning on releasing more of the crates, they will likely be FSL and each of those crates will have a LICENSE.md override. I think this is the canonical pattern but anyone please correct me

bingo! and that's where this journey began, it's been fun proving it out ^_^

the colors are based on the event name.... green for happy path events (READY, DONE, PASS), red for failure events (FAIL, ERROR) and purple for everything else

they don't indicate whether the transition is guarded or not... that is shown in the sidebar when you click a state.

you're not wrong and trimming context is legitimately the first thing that everyone should do. even with context trimming and a tight prompt the model still makes judgement calls about which tools to use and when to stop.

that's fine 90% of the time... the state machine for the other 10% where the model's judgement call costs you an hour of debugging later (confidently fixed wrong, or overzealously) or stops a mostly automated thing because it got stuck on the wrong path.

you're hitting the nail on the head... rules in prompts are suggestions the model can rationalize away.

"the task is so simple that maybe I don't need worktrees" is the model overriding your intent with its own judgement and that's a pattern I'm seeing more and more as these models mature. statewright provides the guardrails... strong suggestions up front on what it can do in X state via injection and if it still wants to try and outsmart that, it gets hit in the post hook and the model gets the message "oh, you're right I shouldn't do it that way" ... instead of you course correcting, it's the state machine

to your first question: the engine is Apache 2.0 and runs locally. the managed service adds the visual editor, run history and plugin install. the enforcement itself doesn't require the cloud, I run the exact same engine on the backend

the MCP server is just the way to get statewright in the hands of a wide array of existing use cases, claude code included. not all agentic clients are created equal and Pi is actually the experience I want to hone next (also the most extensible)

Yes, the engine handles the full workflow schema including guards. There are some aspects of runtime enforcement (env vars/command filtering, etc. exposed via the UI) that currently only live in the plugin layer but the engine parses and exposes everything. All you would have to do is wire up enforcement on your end in your app the same way the plugin does.

Nice project... the per-agent tool restriction is the same core insight (smaller tool space -> better reasoning)

The main difference with Statewright is that tool access changes over time within a single agent. Planning phase gets read-only tools, edit capability unlocks after the agent proves it has adequate understanding... test tools unlock after the fix. State machines handle the phase transitions, guards and retry loops.

Your multi-agent approach decomposes by role instead of by phase/state. Both are valid. Since you're already in Rust, the engine crate (crates/engine) is a pure library with no deps. It might be interesting to see if putting a state machine around your orchestration layer improves your observed performance

Great question... and there are two answers depending on what you were originally referring to:

re: Claude Code... we actually don't filter or modify the tool list so all tools stay visible -- disallowed calls get blocked at execution time with an error message. No cache busts on transitions, the model sees the full tool sets. The cost there is prompt caching dollars not latency I suppose

re: The research (Rust agent + Ollama) the model only receives tool schemas for the current states' allowed tools. Ollama does have a KV cache reuse facility so changing the tool list busts that cache. Depending on your workflow this can happen as many times as you expect your states to transition until completion. For simple workflows this is 3-5x. Within each state the tool list is stable and cache operates normally. Presenting fewer tools instead of dozens on every agent processing step reduces input tokens and decision complexity, which is where the measurable gains come from.

Both enforce the same constraints depending on the execution interface. The schema level filtering in the research is the S-tier approach. Adding tools/list filtering to the MCP gateway would be beneficial if possible (it looks like we could only filter MCP tools not core ones, which could provide tangible benefit. I've added this evaluation to the roadmap.

The workflow definition is intentionally simple... the enforcement layer handles the mechanics however the model gets more context than just "you're in <xyz> mode now"

Each state has an `instructions` field for phase specific guidance and when an agent's action (tool call) gets rejected the error message lets the model know what went wrong, and what's available to move forward

Tool 'Edit' is not available in the 'planning' phase. Allowed Tools: Read, Grep, Glob To advance, call statewright_transition with READY -> implementing

Models (even simple ones) tend to reason through these error messages, adjusting their approaches as opposed to retrying the blocked call. Additionally, on transitions the model is required to include a rationale explaining why it's transitioning (`data.rationale`) which creates an audit trail of the agent's reasoning at each phase boundary. That ends up being one of the most useful parts of the run history viewable on statewright.ai

I also just updated the https://statewright.ai/research page to accurately reflect the intent and mention the patent grant afforded under FSL-1.1-ALv2. Thanks again for calling my attention to this.

Mr. Claude's Opus says that this is a very feasible thing. It has better support for hooks than Cursor and full MCP support so protocol-layer blocking (like Claude) is possible. Adding to the roadmap...

You're right, and I have just corrected this. The license in the repo now uses the canonical FSL-1.1-ALv2 based on the template from fsl.software and now includes the patent grant clause.

The omission wasn't intentional -- the patent grant wasn't on my radar when the original license text was committed. FSL licensing is very new territory for me and I duffed it slightly, now corrected.

Thanks for digging deeper and I'm happy to clarify all three aspects:

Re: Reproducing the results: the engine, agent crate and demo TUI are all in the repo. If you have ollama running with a 13B+ model, task run:bugfix reproduces the simple bugfix result end to end. What isn't published yet is the SWE-bench experiment harness (task selection, patch scoring, control runs). I need to get that out, I prioritized the end-to-end simple Claude Code plugin for the launch. The demo crate (crates/demo) contains a demo TUI which calls ollama and runs the bugfix state machine interactively with code.

Re: Engine: The core engine (crates/engine/) is the pure Rust state machine evaluator. It's what Statewright is running on the backend. JSON in => transition decisions out. Agent (crates/agent/) builds on top of it to make it useful for LLMs. That all is Apache 2.0 with no restrictions.

Re: the Patent: The patent covers the method of using state machines to constrain LLM agent tool access at the protocol layer. It's defensive, it helps protect the managed service and the idea from "being scooped" from a larger company with more personnel and resources. It's not targeted against solo developers, self-hosters or researchers.

You'll find that the portions that I've released FSL 1.1 have explicit grants which do not restrict solo developers or single team self-hosting. The code released this way becomes Apache 2 in exactly 3 years. This is not unlike what Sentry and MariaDB did. I am planning on releasing more portions as FSL 1.1, I just hadn't crossed that bridge and honestly this thing seems to have gotten popular at the moment so I thought I'd set the record straight a bit

agreed... the tooling layer (desktop and console) is where the leverage is right now. the models are good enough, the harness operating them (even us humans) is what's holding things back. that's the basic gist behind this whole project

the LLM doesn't determine the state... it requests a transition to change the state. the engine evaluates guards (data carried along the way) to decide if the transition is valid.

it (the LLM) can't skip from implementation to deploy if the guard says the tests haven't passed. the model will receive feedback that what it's tried to do is invalid and give the reasons why. it can't be skipped. it then tries to resolve that new information to make the state transition... almost like it would responding to a human in the chair denying a step.

the model can't merge if it hasn't gone through your review state, even if it wants to (it'll try though)

the state engine is the part that can't hallucinate. even with simple steps/prompting the review model can miss things... it's still an LLM making a judgement call at the end of the day.

the state engine doesn't judge, it enforces... with code and not transformers ^_^

if a tool (or any other guardrail) isn't valid at a given state the model call gets rejected before the model sees the result. that's the gap between "a model said this is okay" vs. "the system structurally prevents this"

I feel you on how sluggish Claude Code can be, you just never know what those pulsing prompts are doing in the background...

Given Statewright plugs into Claude Code, there is a little added overhead while managing the state machine logic, but for complicated workflows if it saves you a few debug loops, mass edit reversions or death spirals I think the case can be pretty solid for including it

Stately (and XState ^_^) is pretty neat, I hadn't come across it yet... (edit:) neat to see visual XState being used for application logic as well

I see constant posts on Reddit/HN about the ways that AI is amazing and at the same time is fudging it (literally). Nobody can make reliability guarantees on something that's non-deterministic and non-idempotent. Nobody's AI workflow suite of tools can claim this. Prompting gets you closer to the mark but still non-deterministic. Breaking down the problem into chunks with valid transition criterion so that even tiny models can step through them I believe gets us closer to where we want to be semantically

for the integration piece that ties into Claude Code and other places where AI is used most frequently? yes I think it does... we're not fighting context in Opus/Sonnet as much as we are in smaller models and we're only adding about 6 tools here which is a smaller footprint than other MCP exposures. Smaller models have a more direct/tight interface that doesn't bloat the tool space in my experimentation (using the core directly)

Exactly right, and the next step is making it enforceable rather than aspirational. Restrict the plan phase to read-only tools so the agent literally can't edit during planning. Restrict the impl phase to the edit tools the plan identified. Even this basic formalization: same model, same task... yielded a dramatic capability improvement on sub-20B models in my testing.

There's a third mode that works better: structured phases (scoped to each feature, like humans do). (Plan phase => human reviews plan) => (Implement phase => human reviews diff) => (Test phase => tests run). The current TUI tooling gives you the option to do this type of bite-size scoping but you have to enable plan mode and not auto-accept edits. I've been taking the (enhanced) phased approach out of the default toolchains, having discrete phases (even simple plan=>implement=>test) that loop while capping tool access and edit sizes and that's been really promising in the realm of obtaining better agentic coding quality