Is your feature request related to a problem? Please describe.
The README describes OmegaClaw as providing "auditable proof trails for its reasoning," but there is no mechanism to actually capture one. When an iteration misbehaves, there is no linkable record tying together the LLM call, the parsed actions, the policy decisions, the error (if any), and the result — you're left grepping unstructured logs.
Describe the solution you'd like
A per-turn structured reasoning trace: one JSONL record per loop event, linked by a per-iteration trace_id, covering the LLM call (provider/model/latency), action parse, policy decisions, errors, and the final result — plus a small CLI to summarize a trace file (per-iteration LLM latency, parse ok/error counts, policy denials). It should be:
- best-effort (a trace failure never breaks the loop),
- privacy-safe by default (metadata only; redacted bodies only behind an explicit opt-in),
- easy to disable, and configurable in output path.
Describe alternatives you've considered
Parsing the existing INFO logs — but they aren't structured, aren't linked per iteration, and mix third-party noise.
Additional context
- OmegaClaw version: v0.1.17 (current
main)
Is your feature request related to a problem? Please describe.
The README describes OmegaClaw as providing "auditable proof trails for its reasoning," but there is no mechanism to actually capture one. When an iteration misbehaves, there is no linkable record tying together the LLM call, the parsed actions, the policy decisions, the error (if any), and the result — you're left grepping unstructured logs.
Describe the solution you'd like
A per-turn structured reasoning trace: one JSONL record per loop event, linked by a per-iteration
trace_id, covering the LLM call (provider/model/latency), action parse, policy decisions, errors, and the final result — plus a small CLI to summarize a trace file (per-iteration LLM latency, parse ok/error counts, policy denials). It should be:Describe alternatives you've considered
Parsing the existing INFO logs — but they aren't structured, aren't linked per iteration, and mix third-party noise.
Additional context
main)