You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fix(loop): deliver recovered tool-panic message to the LLM + review cleanups (#26)
* fix(loop): deliver recovered tool-panic message to the LLM + review cleanups
A panicking tool's recovered error message was discarded: the goroutine
returned before results[idx] was assigned, so the LLM received an empty
tool result and the consecutive-error tracker never counted the failure.
Drop the early return so the panic message flows through like any other
tool error. The existing TestToolPanic_DoesNotKillAgent never exercised
this path (its message-count trigger never matched with no system
prompt); switch it to a call counter and assert the panic message
reaches the LLM.
Review cleanups from the #25 follow-up:
- loop: name the thrice-repeated anonymous interface as trustAllSetter
- telegram: parse ODEK_TELEGRAM_ALLOW_ALL with strconv.ParseBool for a
consistent truthy grammar (malformed values fail closed)
- telegram: add TelegramConfig.HasAllowlist() and use it in both
ValidateConfig and the open-bot startup warning
- telegram: correct HandlerConfig allowlist comments (AllowAllUsers is
only consulted when BOTH lists are empty) and document on HandleUpdate
that every routed handler must enforce isAllowed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(telegram): pin strict ALLOW_ALL grammar and HasAllowlist behavior
Verification-protocol remediation for #26: the claim that malformed or
extended truthy values ("yes", "on", "junk") fail closed had no
corresponding test, and HasAllowlist had no direct unit test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
0 commit comments