src/inference/runtime_config.py (2)

587-587: stop_sequences not env-overridable — document or add support.

stop_sequences is only sourced from YAML/defaults with no GOVON_GENERATION_STOP_SEQUENCES env var override. This is likely intentional since parsing a list from an env var requires a delimiter convention. Consider adding a brief comment explaining this is YAML-only, or implement env var support with comma-separated parsing if runtime override is needed.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/inference/runtime_config.py` at line 587, The stop_sequences value in
runtime_config.py (constructed from gen_raw and set via
stop_sequences=gen_raw.get("stop_sequences", ["[|endofturn|]"])) is not
overridable via environment variables; either document that stop_sequences is
YAML-only or add env support: read a GOVON_GENERATION_STOP_SEQUENCES env var,
split it on a chosen delimiter (e.g., comma) and strip items, and fall back to
gen_raw if the env var is absent/empty; implement this logic where
stop_sequences is assigned so callers using stop_sequences get the env-parsed
list.

---

498-507: Frozen dataclass with mutable Dict field — works but document the contract.

_ToolsConfig is frozen=True but contains overrides: Dict[str, _ToolDefaultConfig]. While the dict reference is immutable (you can't reassign self.overrides), the dict contents could theoretically be mutated after construction. This is safe here because:

1. The dict is only populated in GovOnConfig.load() and never modified afterward
2. _ToolDefaultConfig is also frozen

Consider adding a brief docstring note that callers must not mutate overrides, or use MappingProxyType for true immutability if this becomes a concern.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/inference/runtime_config.py` around lines 498 - 507, _toolsConfig is
declared frozen but contains a mutable dict `overrides`, which can be mutated
after construction; update _ToolsConfig to enforce immutability or document the
contract: either wrap `overrides` in an immutable view (e.g., use
MappingProxyType on the dict provided in GovOnConfig.load()) so the `overrides`
field truly cannot be mutated, or add a brief docstring on _ToolsConfig stating
callers must not mutate `overrides` after construction; reference symbols:
_ToolsConfig, overrides, _ToolDefaultConfig, GovOnConfig.load(), and for_tool
when making the change.

src/inference/graph/nodes.py (1)

653-656: Minor inconsistency: MAX_TOOL_RESULT_CHARS read inside function vs module-level.

Unlike other constants (lines 35-51) which are read at module import time, MAX_TOOL_RESULT_CHARS is read inside make_tools_node_v3(). This creates a minor inconsistency — the value is captured when the factory is called rather than at import time. While functionally equivalent given the frozen config, consider moving this to module level for consistency:

♻️ Optional: Move to module level for consistency

 _AGENT_INPUT_BUDGET = govon_config.context.agent_input_budget
+_MAX_TOOL_RESULT_CHARS = govon_config.context.max_tool_result_chars

Then reference _MAX_TOOL_RESULT_CHARS in the function.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/inference/graph/nodes.py` around lines 653 - 656, The constant
MAX_TOOL_RESULT_CHARS is being read inside make_tools_node_v3 which is
inconsistent with other module-level constants; move the read to module scope by
defining a module-level variable (e.g., _MAX_TOOL_RESULT_CHARS =
govon_config.context.max_tool_result_chars) near the other constants and update
make_tools_node_v3 to reference _MAX_TOOL_RESULT_CHARS instead of reading
govon_config again.

src/inference/api_server.py (1)

79-87: Verify default_factory usage in non-frozen dataclass.

Using default_factory=lambda: govon_config.generation.max_tokens is valid but creates a closure that reads from govon_config at instance creation time rather than at class definition time. Since govon_config is frozen and immutable after module load, this works correctly. However, if you want the default captured at class definition time (like other modules), consider using a module-level constant:

♻️ Alternative: Use module-level constants for consistency

+_DEFAULT_SAMPLING_MAX_TOKENS = govon_config.generation.max_tokens
+_DEFAULT_SAMPLING_TEMPERATURE = govon_config.generation.temperature

 `@dataclass`
 class SamplingParams:
     """Sampling parameters for vLLM HTTP API. Works without a direct vLLM import."""

-    max_tokens: int = field(default_factory=lambda: govon_config.generation.max_tokens)
-    temperature: float = field(default_factory=lambda: govon_config.generation.temperature)
+    max_tokens: int = _DEFAULT_SAMPLING_MAX_TOKENS
+    temperature: float = _DEFAULT_SAMPLING_TEMPERATURE

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/inference/api_server.py` around lines 79 - 87, The SamplingParams
dataclass currently uses default_factory lambdas that defer reading govon_config
until instance creation; to capture defaults at module-import time for
consistency, define module-level constants (e.g., MAX_TOKENS_DEFAULT,
TEMPERATURE_DEFAULT) assigned from govon_config.generation.* and then use those
constants as the dataclass field defaults (replace field(default_factory=...)
with field(default=MAX_TOKENS_DEFAULT) / field(default=TEMPERATURE_DEFAULT) for
the corresponding fields in SamplingParams) so the values are fixed at
class-definition time.

config/govon.yaml (1)

124-134: Clarify intent: Some tool overrides only specify timeout_sec.

issue_detector, stats_lookup, and demographics_lookup only override timeout_sec without max_retries. According to runtime_config.py (lines 714-717), missing keys fall back to tls_defaults, which means max_retries=0. This is likely intentional, but consider adding a brief comment to make this explicit for operators tuning the config.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@config/govon.yaml` around lines 124 - 134, Add an inline comment in the
overrides block clarifying that omitting max_retries for api entries
(specifically issue_detector, stats_lookup, and demographics_lookup) is
intentional because runtime_config.py falls back to tls_defaults (resulting in
max_retries=0); reference the overrides keys (issue_detector, stats_lookup,
demographics_lookup) and the tls_defaults/max_retries behavior so operators know
this omission is deliberate when tuning timeouts.