Dannon Baker deb7ecd706 Make agent retry budget configurable via inference_services ...

ChatGXY's router was hard-failing with 'Exceeded maximum output retries (1)':
its pydantic-ai Agent used the default retry budget of 1, and the router's
process() only caught (OSError, ValueError), so an UnexpectedModelBehavior sailed
past the graceful _handle_fallback and surfaced to the user as a generic 500.

This adds a per-agent 'retries' knob to inference_services, resolved through the
same _get_agent_config chain as temperature/max_tokens and wired into every
agent's Agent(retries=...). The default goes from pydantic-ai's 1 to 3 so a
transient bout of non-conforming model output recovers instead of failing the
run. custom_tool's producer keeps its retries=0 reflection-loop default but can
now be overridden. The router also catches UnexpectedModelBehavior so a genuinely
exhausted budget degrades to the existing fallback message instead of a 500.

Adds unit tests for the resolution chain, the config->Agent wiring, and the
router's graceful fallback (forced via an empty ModelResponse).

2026-06-12 12:59:27 -04:00

..

relicensing

Audit of pre-2021-04-27 contributors to check for relicensing

2026-03-03 16:35:53 +00:00

source

Make agent retry budget configurable via inference_services

2026-06-12 12:59:27 -04:00

Makefile

Fix Makefile rule and update `run_tests_help.txt`

2022-09-09 17:20:23 +01:00

parse_gx_xsd.py

Apply 2026 black style

2026-01-20 10:25:45 +00:00

schema_template.md

Add credentials tags to schema template for tool requirements

2026-01-12 17:41:17 +01:00