Local code execution

Code repair, with checks attached.

Sobek Stream runs bounded engineering tasks against local source. A Worker proposes a typed change; compilers, tests, and policy decide whether it stays.

Follow a task ↓ Read the API ↗

source stays local models GPU only record source-free

01 / RUNTIME

Declare the boundary. Run inside it.

Planning, mutation, and acceptance are separate steps. A candidate becomes a result only after the declared checks agree.

01 / request

State the work

A workflow, task contract, write scope, and retention policy enter through the API.

02 / context

Expose less

Workerd gives the execution boundary an opaque local workspace and a constrained capsule.

03 / action

Make one change

GPU Workers return typed transactions. Every operation is checked against the declared scope.

04 / proof

Keep what passes

Compilers, tests, Breakers, and Judges leave a replayable Ledger record.

02 / WORKFLOWS

Different work. One contract.

Seven versioned workflows share the same isolation, scheduling, provenance, and evidence boundary.

Matching decompilation

Bounded source transforms toward exact object equivalence.

decomp.matching.v1 ↗ 02

Compiler repair

Small fixes checked by the declared compiler and scope.

repo.compiler_repair.v1 ↗ 03

Repository bugfix

Isolated changes followed by objective validation.

repo.bugfix.v1 ↗ 04

Test generation

Regression coverage with executable acceptance Gates.

repo.test_generation.v1 ↗ 05

API migration

Scope-sensitive replacements with a retained decision trail.

repo.api_migration.v1 ↗ 06

Refactoring

Typed rename and multi-file operations without blind merge.

repo.refactor.v1 ↗ 07

Performance work

Keep a change only when the benchmark Gate says it improved.

repo.performance.v1 ↗

03 / MODELS

Measure the Worker, then decide.

The L40S comparison now spans the initial small roster and four recent 35B-A3B tracks. These are qualification signals, not general capability claims.

Qwen3.6 35B-A3B

Semantic-repair challenger · 12/12 Worker-sourced · 9/9 safe abstentions

Q4_K_M · 1,658 ms p50

Qwen-AgentWorld 35B-A3B

Specialist candidate · 15/15 Worker-sourced · separate from general coding

Q4_K_M · 1,186 ms p50

Qwen3.5 35B-A3B Base

Worker candidate · 15/15 Worker-sourced · community Q8_0 warning

Q8_0 · 1,905 ms p50

Qwen3.5 35B-A3B

Post-trained candidate · 9/9 objective · four deterministic fallbacks

Q4_K_M · preliminary

GPT-OSS 20B

Frontier control · 4/12 Worker-sourced on the matched semantic lane

MXFP4 · control

Qwen2.5-Coder 0.5B Base

Operational small-Worker default · 5/5 matched non-toy accepts

Q4_K_M · default

LFM2.5-350M Base

Latency challenger · 3/5 matched non-toy accepts, 2 safe fallbacks

Q4_K_M · challenger

IQ4_XS deployment variant

Qwen3.6 challenger · 0/15 Worker-sourced · safe but rejected

18.5 GB · rejected

Qwen3.6 exceeds GPT-OSS 20B on the matched semantic-repair lane at 12/12 versus 4/12 Worker-sourced positives, with safe negative controls for both. The four 35B-A3B tracks and the small-Worker comparison are profile-scoped evidence, not a global release claim: broader comparative quality and cross-hardware reproduction remain open. The IQ4_XS variant is safe but rejected because its positive accepts used deterministic fallback. See the model evaluation plan ↗ and benchmark record ↗ for hashes, controls, and the qualification boundary.

04 / EVIDENCE

Proof belongs beside the result.

Discovery, task events, provenance, and retained evidence are API resources. Source and raw model output stay out of the retained record.

request shapestream / 1.0

POST /v1/stream/tasks

{
  "task_version": "1.0",
  "task_id": "rename-widget",
  "workflow": "repo.refactor.v1",
  "workspace": { "workerd": "local", "path_ref": "/approved/example-repo" },
  "goal": "Rename Widget without changing behavior",
  "acceptance": { "write_scope": ["src/widget.rs"] },
  "policy": { "retention": "metadata_only" }
}

202 Accepted  →  Gates  →  Ledger

current qualificationrelease gates open

Stream API + Workerdimplemented
GPU inference policyCPU disabled
Measured L40S native routeGPU-only
Qwen3.6 semantic repair12 / 12 sourced
Qwen3.6 negative control9 / 9 abstained
Four 35B-A3B tracksscreened
Qwen3.6 vs GPT-OSS semantic12 / 12 vs 4 / 12
Matched small-Worker controlsQwen 5 / 5 · LFM 3 / 5
IQ4_XS variantrejected
K2 Horizon safety screen9 / 9 abstained
K2 → Qwen Judge integration3 / 3 accepted
Comparative release gateopen
Cross-hardware releasenot yet
External proxy checkpassed

Qwen3.6 now clears the GPT-OSS 20B control on the semantic-repair Worker lane. These are still profile-scoped controls, not a release claim: broader comparative quality, Judge/Captain qualification, cross-hardware reproduction, and variant role probes remain open.

Open dashboard JSON ↗