diff --git a/.opencode/agents/reviewer.md b/.opencode/agents/reviewer.md
new file mode 100644
index 0000000000..26ea3950d3
--- /dev/null
+++ b/.opencode/agents/reviewer.md
@@ -0,0 +1,40 @@
+---
+description: Read-only code reviewer for pull requests
+mode: primary
+model: cf-gateway/kimi-k2.5
+temperature: 0.1
+permission:
+  edit: deny
+  bash:
+    "*": deny
+    "git diff*": allow
+    "git log*": allow
+    "git show*": allow
+  webfetch: deny
+---
+
+You are a code reviewer for the Cog project. Cog packages ML models into
+production-ready containers. The codebase has three languages: Go (CLI and
+tooling), Python (SDK), and Rust (coglet prediction server).
+
+## How to review
+
+1. Read the full diff. Understand what the PR is trying to do before commenting.
+2. Load the appropriate review skill(s) for the languages touched in the PR:
+   - Go files: load `go-review`
+   - Python files: load `python-review`
+   - Rust files: load `rust-review`
+   - Architecture/cross-cutting: load `cog-review`
+3. Focus on substance, not style. Linters handle formatting.
+4. Scope feedback to the PR's purpose. Don't demand unrelated refactors.
+5. Be direct. Say what the problem is, why it matters, and how to fix it.
+
+## Comment format
+
+- Reference specific files and lines: `path/to/file.go:42`
+- Use ```suggestion blocks when proposing concrete code changes
+- Categorize findings:
+  - **Blocker**: Must fix before merge (bugs, security, data loss)
+  - **Should fix**: Important but not blocking (error handling gaps, edge cases)
+  - **Nit**: Minor improvement, take it or leave it
+- If the PR looks good, say so briefly. Don't invent problems.
diff --git a/.opencode/skills/cog-review/SKILL.md b/.opencode/skills/cog-review/SKILL.md
new file mode 100644
index 0000000000..fbd66d0ea9
--- /dev/null
+++ b/.opencode/skills/cog-review/SKILL.md
@@ -0,0 +1,51 @@
+---
+name: cog-review
+description: Cog architecture and cross-cutting review guidelines
+---
+
+## Cog architecture review guidelines
+
+Cog packages ML models into production-ready containers. Use this skill for
+changes that cross language boundaries or touch core architecture.
+
+### Component overview
+
+- **CLI** (Go): `cmd/cog/` and `pkg/` -- builds, runs, and deploys models
+- **Python SDK**: `python/cog/` -- predictor interface, types, HTTP/queue server
+- **Coglet** (Rust): `crates/` -- prediction server inside containers (HTTP, worker management, IPC)
+
+### Key design patterns
+
+**Wheel resolution**: The CLI discovers SDK and coglet wheels from `dist/` at
+Docker build time. Wheels are NOT embedded in the binary. Changes to build
+artifacts need to account for this.
+
+**Dockerfile generation**: `pkg/dockerfile/` generates Dockerfiles from `cog.yaml`
+config. Template injection and escaping matter here.
+
+**Config parsing**: `pkg/config/config.go` parses `cog.yaml`. Schema is at
+`pkg/config/data/config_schema_v1.0.json`. Changes must keep schema and Go
+code in sync.
+
+**Two-process coglet**: Parent process (HTTP server + orchestrator) and child
+worker process (Python predictor execution) communicate via IPC. Changes to
+the IPC protocol affect both Rust and Python code.
+
+**Compatibility matrix**: CUDA/PyTorch/TensorFlow compatibility is managed in
+`tools/compatgen/`. Framework version changes have wide blast radius.
+
+### Cross-cutting concerns
+
+- **VERSION.txt** is the single source of truth for versioning. `Cargo.toml` must match.
+- Changes to `python/cog/base_predictor.py` affect all downstream model authors.
+- Changes to `docs/` require running `mise run docs:llm` to regenerate `docs/llms.txt`.
+- CLI reference docs are auto-generated -- changes to `cmd/` or `pkg/cli/` require `mise run docs:cli`.
+- Integration tests in `integration-tests/` use Go's testscript and need a built cog binary.
+
+### What to watch for
+
+- Breaking changes to the predictor interface or cog.yaml schema
+- Docker build changes that affect layer caching or build time
+- IPC protocol changes without updating both Rust and Python sides
+- Version bumps that miss one of the places VERSION.txt needs to match
+- New ML framework versions without compatibility testing
diff --git a/.opencode/skills/go-review/SKILL.md b/.opencode/skills/go-review/SKILL.md
new file mode 100644
index 0000000000..60c5a43558
--- /dev/null
+++ b/.opencode/skills/go-review/SKILL.md
@@ -0,0 +1,44 @@
+---
+name: go-review
+description: Go code review guidelines for the Cog codebase
+---
+
+## Go review guidelines
+
+This project uses Go for the CLI (`cmd/cog/`, `pkg/`) and support tooling (`tools/`).
+
+### What linters already catch (skip these)
+
+golangci-lint runs errcheck, gocritic, gosec, govet, ineffassign, misspell,
+revive, staticcheck, and unused. Don't flag issues these would catch.
+
+### What to look for
+
+**Error handling**
+- Errors returned but not checked or silently discarded
+- User-facing errors should use `pkg/errors.CodedError` with error codes
+- Generic error wrapping that loses context (`fmt.Errorf("failed")` with no `%w`)
+
+**Imports**
+- Should be three groups: stdlib, third-party, internal (`github.com/replicate/cog/pkg/...`)
+- Only flag if actually wrong, not cosmetic reordering
+
+**Testing**
+- Must use `testify/require` for fatal assertions and `testify/assert` for non-fatal
+- No raw `if` checks with `t.Fatal`/`t.Errorf`
+- Prefer table-driven tests for similar cases
+- Prefer specific assertions (`Equal`, `Contains`, `NoError`) over `True`/`False`
+
+**Concurrency**
+- Goroutine leaks (no cleanup path, missing context cancellation)
+- Shared state without synchronization
+- Channel misuse (sends on closed channels, unbuffered channels in wrong contexts)
+
+**Docker/container patterns**
+- The CLI uses the Docker Go SDK. Watch for leaked clients, unclosed response bodies
+- Dockerfile generation is in `pkg/dockerfile/` -- template injection risks
+
+**Architecture**
+- Commands belong in `pkg/cli/`, business logic in `pkg/`
+- Config parsing/validation in `pkg/config/`
+- Don't mix CLI concerns with library logic
diff --git a/.opencode/skills/python-review/SKILL.md b/.opencode/skills/python-review/SKILL.md
new file mode 100644
index 0000000000..7ac56147ce
--- /dev/null
+++ b/.opencode/skills/python-review/SKILL.md
@@ -0,0 +1,47 @@
+---
+name: python-review
+description: Python code review guidelines for the Cog SDK
+---
+
+## Python review guidelines
+
+This project uses Python for the SDK (`python/cog/`) which defines the predictor
+interface, type system, and HTTP/queue server.
+
+### What linters already catch (skip these)
+
+ruff handles pycodestyle (E), Pyflakes (F), isort (I), warnings (W), bandit (S),
+bugbear (B), and annotations (ANN). Don't flag issues these would catch.
+
+### What to look for
+
+**Type annotations**
+- Required on all function signatures
+- Use `typing_extensions` for backward compatibility
+- Avoid `Any` where a concrete type is possible
+- Check that type annotations actually match runtime behavior
+
+**Compatibility**
+- Must support Python 3.10 through 3.13
+- Watch for syntax/stdlib features only available in newer versions
+- `from __future__ import annotations` if using newer annotation syntax
+
+**Error handling**
+- No bare `except:` or `except Exception:` that swallows everything
+- Exceptions should have descriptive messages
+- Resource cleanup with context managers, not try/finally when avoidable
+
+**Async patterns**
+- Tests use pytest-asyncio -- async tests need proper fixtures
+- Watch for blocking calls inside async functions
+- Proper cleanup of async resources (aclose, async context managers)
+
+**Predictor interface**
+- `base_predictor.py` is the core interface -- changes here affect all users
+- `types.py` defines input/output types -- check backward compatibility
+- Server code in `python/cog/server/` handles HTTP -- watch for request handling bugs
+
+**Testing**
+- Uses pytest with fixtures
+- tox runs tests across Python 3.10-3.13
+- Test isolation: don't rely on global state or test ordering
diff --git a/.opencode/skills/rust-review/SKILL.md b/.opencode/skills/rust-review/SKILL.md
new file mode 100644
index 0000000000..04f8881857
--- /dev/null
+++ b/.opencode/skills/rust-review/SKILL.md
@@ -0,0 +1,49 @@
+---
+name: rust-review
+description: Rust code review guidelines for Coglet
+---
+
+## Rust review guidelines
+
+This project uses Rust for Coglet (`crates/`), the prediction server that runs
+inside Cog containers. It handles HTTP requests, worker process management, and
+prediction execution.
+
+### What linters already catch (skip these)
+
+clippy runs in CI. cargo-deny audits dependencies for license/advisory issues.
+Don't flag issues these would catch.
+
+### What to look for
+
+**Error handling**
+- Use `thiserror` for typed errors in library code, `anyhow` for application errors
+- Don't use `.unwrap()` or `.expect()` in non-test code unless the invariant is documented
+- Error context: use `.context()` or `.with_context()` from anyhow, not bare `?`
+
+**Ownership and lifetimes**
+- Unnecessary cloning where a borrow would work
+- Lifetime issues that suggest a design problem (not just annotation noise)
+- Arc/Mutex when simpler patterns exist
+
+**Async (tokio)**
+- Blocking operations inside async contexts (use `spawn_blocking`)
+- Missing `.await` on futures (compiler catches some, but not all logical issues)
+- Proper cancellation handling and cleanup
+- Task/JoinHandle leaks
+
+**Safety**
+- Any `unsafe` block needs justification and a safety comment
+- FFI boundaries with Python (PyO3) -- check for panics across FFI, GIL handling
+- Memory safety in IPC between parent and worker processes
+
+**Architecture**
+- `crates/coglet/` is the core: HTTP server, worker orchestration, IPC
+- `crates/coglet-python/` is PyO3 bindings for Python predictor integration
+- Two-process architecture: parent (HTTP + orchestrator) and worker (Python execution)
+- Don't mix IPC concerns with HTTP handling
+- Snapshot tests use `insta` -- check if snapshots need updating
+
+**Dependencies**
+- New dependencies should be justified -- `crates/deny.toml` audits them
+- Prefer std or existing deps over adding new ones for small tasks