v0.5.4

What's Changed

New Features

• Self-update command (#111) — zeptoclaw update checks GitHub Releases for the latest version, downloads the matching platform binary, verifies its SHA256 checksum, and atomically replaces the running executable

  zeptoclaw update              # update to latest
  zeptoclaw update --check      # check without downloading
  zeptoclaw update --version v0.5.2  # specific version
  zeptoclaw update --force      # re-download even if current
  

Bug Fixes

• OpenAI token field — o1, o3, o4, and gpt-5 models require max_completion_tokens instead of max_tokens. Previously the provider sent the wrong field, received a rejection, then retried — wasting a round-trip on every agent call. Now selects the correct field upfront

• Tool execution race (#113) — When the LLM called write_file and read_file in the same batch, parallel execution caused a read-before-write race. Tool batches containing filesystem writers now execute sequentially in LLM-declared order; read-only batches remain parallel

• Scheduling & delivery (#113) — Four bugs in the cron/heartbeat pipeline:

  • System prompt now instructs the LLM to deliver Reminder: X messages as friendly notifications and respond HEARTBEAT_OK to heartbeat pings
  • One-shot cron job retain logic rewritten with a clear should_remove variable (was a confusing double-negative)
  • ZEPTOCLAW_HEARTBEAT_DELIVER_TO=channel:chat_id config now actually wires through to HeartbeatService
  • Unknown outbound channel logs downgraded from warn to debug (eliminates noisy spam on every heartbeat tick)

Full Changelog: v0.5.3...v0.5.4