Show HN: Recast, an experimental Smalltalk app that repairs itself with an LLM

A Smalltalk application that uses an LLM to repair itself while it runs.

Recast wires an LLM into a running application's exception path. A failing call supplies the stack, method source, receiver state, and tests. The model can propose a repair; a separate VM checks it before the application replaces the method in its live image. Subsequent calls use the new code, without a rebuild or restart.

The working demo is a vending machine with a deliberately broken dispense method. With automatic repair enabled, an ordinary call triggers the whole loop—no lk heal command or manually written repair prompt. Here is a shortened transcript, after injecting the failure:

$ ./lk eval "| vm | vm := VendingMachine make. vm insertCoin: 60. vm dispense"
ERROR: Error: automatic repair demo

$ ./lk incidents
... "status": "awaiting-model" ...
... "status": "verifying" ...
... "status": "repaired", "detail": { "ok": true, "repro": "true", ... }

$ ./lk eval "| vm | vm := VendingMachine make. vm insertCoin: 60. vm dispense"
'product'

The first call still fails. Investigation and verification run asynchronously; the image applies the accepted patch, and the next call succeeds. This path has been exercised with a real LLM on Pharo 12.

Take the tour in GUIDE.md to reproduce it: boot, edit live code, break the app, trigger automatic repair, and promote the result. You can also use the offline fake LLM without an API key; it supplies canned proposals through the same verification path.

An application's runtime has useful evidence about a failure: the call stack, the receiver's state, and the code that actually ran. Recast turns that evidence into a repair request and connects the result to a live deployment mechanism. The experiment is how much of an application's maintenance loop can become part of the application itself.

Smalltalk is a useful substrate because classes, methods, and execution contexts are objects the running system can inspect. Compile a replacement method into a class, and subsequent message sends use it immediately. The existing application objects keep their identity; applying a method patch does not require rebuilding their world.

Recast adds a verification and persistence policy around that capability: test the proposal in another VM, apply it tentatively, and explicitly promote the definitions that should survive restart. The LLM proposes; the runtime checks and applies.

The example is intentionally small: one demo class, method replacements, receivers with simple literal state, and two application tests. It demonstrates the mechanics of automatic repair, not general autonomous software maintenance. A green test run establishes only what those tests check.

flowchart LR
    Failure[Application call fails] --> Capture[Capture incident]
    Capture --> LLM[LLM investigates]
    LLM --> Decision{Repair justified?}
    Decision -->|Yes| Shadow[Separate VM: repro and tests]
    Decision -->|No| Audit[Record decision]
    Shadow -->|Pass| Live[Hot-swap live method]
    Shadow -->|Fail| Audit
    Live --> Promote[Explicit promotion]

1. Capture.LLMRepairService observes an error at the application's eval boundary, before it becomes an error string. For an opted-in application class, it captures the failing method, stack, source, simple receiver state, arguments, existing tests, and the class's repair contract. It constructs a repro using a fresh receiver with the captured state.
2. Investigate. The Python supervisor sends that evidence to the model. The model returnsexpected ,insufficient-evidence , orrepair , with a reason. A raised exception alone does not require a patch.
3. Verify. The image parses a repair proposal and permits only a replacement of the captured application method. A separate Pharo VM loads the kernel, replays the journal, applies the candidate, and runs the repro and suite.
4. Apply. A passing verdict lets the live image compile the replacement and journal it. The failed operation is never automatically retried. Promotion remains a separate operator action.

The application queue and heartbeat continue while the model and shadow VM work. Repeated failures from an unchanged method are deduplicated. A code edit invalidates an outstanding proposal; disabling repair or restarting cancels pending work. lk incidents records decisions and verification results, and data/heal/req-N.json / resp-N.json preserve the captured evidence and replies.

Automatic repair defaults to off and currently opts in VendingMachine. Expected insufficient-credit errors, compilation errors, test runs, boot replay, and shadow evaluation do not start investigations. Another application class can opt in through repairClasses in kernel configuration and a repairContract class method; repairExpectedErrors lists error messages to exclude.

You need Git, Docker with the Compose plugin, Python 3, and an API key for an OpenAI-compatible Chat Completions endpoint. From a fresh checkout:

git clone https://github.com/rot13maxi/recast.git
cd recast
cp .env.example .env
docker build -f Dockerfile.base -t live-smalltalk-base:latest .
docker compose up -d --build
./lk status
./lk tests

The template includes an OpenAI example endpoint and model. You can configure another compatible provider or a local model. The GUIDE explains the request format and includes a connection check from inside the container. Real-model repair runs have been validated; check your chosen endpoint before starting the tour. If .env already exists, edit it instead of overwriting it. It is ignored by Git. The first build downloads Pharo and its dependencies.

Continue with the GUIDE for the failure-to-repair experiment. The command-line interface also exposes each part of the loop:

For explicit lk heal, the model receives the supplied problem and repro without the automatic incident's source and state capture. Include the relevant fields and intended behavior in the problem description. This path is an optional experiment in the GUIDE.

The live image never saves a heap snapshot. Each boot loads the base image, installs the kernel, replays the promoted journal prefix, and starts the queue. Mutable files live in /data, bind-mounted from the host's data/ directory:

Definitions beyond the marker are tentative. Restart leaves them unapplied; lk promote runs the suite and advances the marker to the journal size if it passes. lk rollback <n> chooses an earlier prefix for the next boot. The journal restores definitions, not application data or in-memory objects.

Real-model runs exercised both automatic and explicit repair. The proposed vending-machine fixes passed the shadow repro and application suite and were applied live. Promotion and restart preserved a repair; an unpromoted edit was left unapplied after restart.

Deterministic integration checks cover rejected candidates, patches outside the allowed method, stale proposals, duplicate suppression, expected failures, declined investigations, and cancellation. A deliberately slow shadow run leaves the application queue and heartbeat responsive. The offline endpoint returns predefined patches; those checks test the mechanism, not model quality. There is no repair success-rate benchmark here.

To rerun the checks against a demo stack:

python3 -m unittest discover -s test -p 'test_*.py'
docker compose stop supervisor
python3 test/auto_repair.py
docker compose start supervisor

The integration script makes tentative edits, restores the vending method and automatic-repair setting, and leaves its journal entries and audit trail.

• Verification depends on tests. The demo has two application tests. The model gets an application contract, but that prose is not an independent executable verifier. Stronger tests and independent specifications matter.
• The shadow VM is not a sandbox. Candidate code can access shared files and runtime facilities. Restricting which method can be replaced does not constrain everything its body can do. Hostile proposals need stronger process and filesystem isolation.
• Repair does not undo side effects. Receiver state is captured at failure time. The failed call may already have changed state, so it is never blindly retried. Complex object graphs are skipped.
• Method repair is narrower than state migration. Changing instance-variable layouts or assumptions about existing objects needs migration machinery. Passing tests in a fresh VM does not establish the live heap's compatibility.
• Replay is code persistence. Application data needs separate storage.

The next experiments are stronger isolation, independent tests, canary traffic, and explicit support for state migration and operations that can be retried.

When extending the kernel, edit the Python sources and run python3 kernel/gen_methods.py. The installer creates classes before compiling methods; method sources are stored as JSON strings to avoid manual quote escaping. Boot and shadow replay avoid live-side effects such as incident capture and automatic test registration.