Freegle / Iznik / 27212

72%

push

circleci

fix(monitor-fsm): recover the brain's JSON decision from prose preambles

The FSM brain (a reasoning model, currently claude-opus-4-8) frequently wraps
its JSON decision in natural language - "I need to...", "The FSM ac..." - which
made ai-flower's JSON.parse throw "Unexpected token 'I'". sanitizeLLMDecision
only stripped markdown fences, so on that failure it returned the raw prose
unchanged, validation failed after 3 retries, and the iteration force-advanced
to COVERAGE_GATE with zero PRs. Two consecutive live iterations produced no
fixes for exactly this reason.

Extract the JSON out of the prose: pull the first balanced JSON object/array
(ignoring braces inside strings) and, when one is found, return it instead of
the raw text. Moved the repair into a testable module (src/llm-json.ts) with
unit tests covering the two real preambles observed in the logs.

13037 of 17188 branches covered (75.85%)

Branch coverage included in aggregate %.

136449 of 189539 relevant lines covered (71.99%)

37.87 hits per line