The README’s example request body uses {"model_id": "..."}, but the handler expects a model field. Either update the handler to accept model_id (or both for compatibility) or fix the README example so users can successfully switch local models.

This channel docstring advertises a configurable control_port, but the implementation ignores config and always binds to 9090. Either add control_port (and test_users, if intended) to the channel config schema and read it here, or remove it from the docs to avoid a misleading configuration surface.

This new test module does not compile against the current config/auth APIs: PolyConfig.proxy is an Option<ProxyConfig> (so config.proxy.default_policy won’t compile), check_model_access returns bool (not Result), and model_matches is private to proxy::auth. Update the tests to use ProxyConfig directly (like proxy/auth.rs tests) or expose the necessary APIs in proxy::auth.

ContextStore has no get_recent(..) method (and this file currently builds a combined string but then discards it), so this won’t compile and delegation won’t include recent context as intended. Either add the needed API to the context store / unified context abstraction, or change this to use the existing context augmentation/preamble mechanism.

/v1/chat/completions trusts the caller-provided x-agent-id header but does not enforce proxy.api_key (or any agent-specific API key) for authentication. As a result, a client can impersonate any agent and bypass model access controls. Add an Authorization check (global key and/or agent-specific keys) and derive agent_id from the authenticated key rather than from an untrusted header.

unwrap_or(&vec![]) takes a reference to a temporary Vec, which will not compile (borrowed value does not live long enough). Use unwrap_or(&[]) (slice) or bind a local empty Vec/&[] fallback before iterating.

unwrap_or(&vec![]) takes a reference to a temporary Vec, which will not compile. Use unwrap_or(&[]) / unwrap_or_default() on an owned collection, or store an empty Vec in a local variable and reference that.

.unwrap_or(&vec![]) references a temporary Vec, which will not compile (borrowed value does not live long enough). Use a slice fallback (unwrap_or(&[])) or bind an empty Vec in a local variable before iterating.

/send doesn’t route the message through the real system yet (it just logs and returns a hard-coded "Mock response"). That makes the mock channel unsuitable for integration testing as described; either implement dispatch via router/command_handler/context_store, or clearly scope this as a stub.

.unwrap_or(&vec![]) again borrows a temporary Vec and won’t compile. Prefer unwrap_or(&[]) / a local empty collection to avoid referencing a temporary.

Under cfg(all(test, loom)) these tests call registry.clone(), but ProviderRegistry does not implement Clone, so this won’t compile. Wrap ProviderRegistry in crate::sync::Arc (or implement Clone) for sharing across loom threads.

This module is not referenced from proxy/mod.rs (no mod retry_gateway;), so it will never be compiled. As written, it also imports symbols that don’t exist in crate::proxy::gateway (e.g., Gateway, BackendError), so it’s likely stale/incomplete—either wire it up properly or remove it to avoid bitrot.

This file isn’t referenced by the module tree (no mod auth_test;), so it won’t be compiled or run. Consider moving these tests to crates/zeroclawed/tests/ (integration tests) or including the module from proxy/mod.rs under #[cfg(test)] so auth behavior is actually exercised.

The README example uses model patterns like "llama*", "qwen*", and "gemma*", but the routing implementation only supports exact match, *, and prefix/* globs. As written, these patterns will never match and routing will fall back to the default backend. Either update the docs to use supported patterns (e.g. qwen/*) or extend model_matches_pattern to support * suffix wildcards.

The [local_models] README example doesn’t match the actual config schema: [[local_models.models]] expects hf_id (and optional provider_type/display_name), while port/host are configured under [local_models.mlx_lm]. The example currently uses path, backend, and per-model port, which won’t deserialize. Please align the documentation with LocalModelsConfig/LocalModelDef.

This test module doesn’t currently test caching behavior: test_caching_behavior contains no assertions and just prints a message. Either convert it into a real test using a mock HTTP server (e.g. mockito) to assert cache hits/misses, or remove/mark it clearly as a placeholder so it doesn’t give a false sense of coverage.

crates/zeroclawed/src/proxy/auth_test.rs appears to be an orphaned test file: it isn’t referenced from the module tree, and it uses an old check_model_access(...).is_ok()/is_err() API that doesn’t match the current bool return type. Consider either wiring it in via #[cfg(test)] mod auth_test; (after updating it to compile) or deleting it to avoid confusion and bitrot.

This handler builds a context object with cwd/session_id, but it isn’t passed into PolicyEngine::evaluate (the engine constructs its own context), so cwd will always be empty in Starlark policies. Either remove the unused context construction, or extend PolicyEngine::evaluate to accept and merge additional context fields so policies like claude-code-policy.star can actually use cwd/session_id.

The Loom tests call registry.clone(), but ProviderRegistry doesn’t implement Clone, so this will not compile under cfg(loom). Consider wrapping the registry in crate::sync::Arc inside the Loom test (and cloning the Arc), or implement Clone for ProviderRegistry by storing the inner map behind an Arc<RwLock<...>>.

run_hook spawns an external process and waits for it to finish, but there’s no timeout/kill logic. A hung hook will stall the request indefinitely even though the comment implies bounded time. Consider enforcing a timeout (e.g., tokio::time::timeout around spawn_blocking and/or killing the child after a deadline) so voice requests can fail fast or degrade gracefully.

A new reqwest::Client is constructed for every STT/TTS request. Creating clients is relatively expensive (connection pools, TLS config) and can become a throughput bottleneck. Consider reusing a single client (e.g., store it in ProxyState or cache per-upstream config) rather than rebuilding per call.

The cargo build pipeline masks build failures (... | grep ... || true) even though set -euo pipefail is enabled. If the build fails, the script will still print "Build complete" and then cp non-existent binaries, producing confusing errors. Consider removing the || true (and/or using grep ... || { cat; exit 1; }) so the script aborts on build failure while still filtering output.

UnifiedContextStore::new(...) can return a persistent store when persistent-context is enabled and [context.persistent] is configured, but main immediately calls into_in_memory()?, which will bail for the persistent variant. This makes the persistent-context feature unusable in practice. Consider plumbing UnifiedContextStore (or a dyn ContextStoreTrait) through channels/CommandHandler instead of forcing an in-memory ContextStore clone.