[fchengyan (fcy222fcy)]

Summary

Fixes #131. With "verbose": true in the user's Claude Code settings, every Claude Code side call failed, and the micro-skills proposer's unguarded runner call aborted the whole fold.

Two defects, both in the #131 report:

1. claude_call mis-parsed the reply. Under -p --output-format json, a verbose Claude Code prints a JSON array of stream events (system, assistant, ..., result), not the bare result object. json.loads then data.get("result") raised AttributeError: 'list' object has no attribute 'get'.
2. Distiller._maybe_propose called its runner unguarded. Unlike the loop guard's judge (loopguard/judge.py catches runner exceptions and records runner-error:...), the proposer let the exception escape, so the fold failed: the session stayed pending, got no ledger row, and was re-folded — with another paid Haiku call — on a later eligible SessionStart event.

What changed

• tokeneconomizer/plugin/sidecall.py — new _claude_json_result(stdout) normalizes what the CLI printed: a list's last type == "result" event, the bare object as before, or {}. claude_call and nothing else changes shape: the usage/modelUsage accounting and the _no_reply path are untouched.
• tokeneconomizer/microskills/distiller.py — _maybe_propose guards its runner like the judge does: a failed side call skips the proposal instead of aborting the fold, and backs off one cadence (last_propose = now) so a broken runner is not paid again on every tick.
• tests/test_sidecall_verbose.py — three regression tests (below).

I went with shape-accepting rather than the report's --settings '{"verbose":false}' suggestion: it needs no capability probe of --settings, and it keeps working for a CLI that prints the array for any other reason.

Verification

Red/green with the subprocess faked the way test_live_core.py fakes it (the reporter's real-CLI repro is not available on this machine):

• verbose-array reply with a result event → before: AttributeError: 'list' object has no attribute 'get' at sidecall.py:235; after: returns ("OK", eff) with usage charged.
• verbose-array reply with no result event → before: same AttributeError; after: comes back empty through the _no_reply path.
• a runner that raises in _maybe_propose → before: the exception escaped and aborted the fold; after: the proposal is skipped and last_propose backs off.

Reverse control: git stash of the two source files (tests kept) turns all three red; restoring turns them green. Full offline suite (python -m pytest tests -q) is green on Python 3.11 / Windows, including test_live_core.py and test_live_repos.py.

AI disclosure

This fix was authored with AI assistance (Claude agent, at the direction of fcy222fcy).

[fchengyan (fcy222fcy)]

@microsoft-github-policy-service agree

[SakshamUboweja]

Thanks for this, and for the clear red/green write-up. #134 fixes #131 the same way and also logs the failed proposer call to errors.log, so I’m merging that one and closing this as a duplicate. Your handling of a non-object JSON reply (e.g. null) is a nice touch that #134 doesn’t have. A small follow-up PR on top of #134 for that case would be welcome.