Skip to content

v0.4.1

Choose a tag to compare

@github-actions github-actions released this 19 Jun 01:46
· 28 commits to main since this release

Changed

  • Power through to auto-compaction — stop getting "scared" near the context ceiling.
    Field report: agents were stopping/handing back control as context filled, which prevents
    the very auto-compaction that would refresh them and continue the task. The skill now
    states it outright: stopping near the ceiling strands the task; keep producing work
    until Claude Code auto-compacts, then resume from the handoff. Removed "context at the
    compaction floor" from the legitimate stop-conditions list — only quota (rate-limit)
    ever justifies stopping; context never does.
  • Non-redundant handoff cadence (throughput). The mid-turn ctx logic no longer nags for
    a handoff once one was saved recently (kills the redundant 6%→3% re-saving). Cadence is
    now: refresh at task boundaries, plus one velocity-timed "super close to
    auto-compaction" nudge (fires once, bypasses the throttle, suppressed if already saved).
    Updates surface "handoff already saved Nm ago" and a "≈N tool calls at this pace" estimate
    (from per-tool-call context growth) so the agent knows it's captured and how close it is.
  • Context handoff-nudge held later — utilize context to the core. ctx_bands lowered so the
    "context getting low" nudge fires near the ceiling (~4–8% left) instead of at 25%/40%. The
    handoff is one cheap call that only needs to land before compaction, and the velocity-timed
    "super close" message remains the final safety net — so the agent uses nearly all its context
    before prepping a handoff, rather than slowing down early.

Added

  • Current time + timezone awareness. Every prompt stamp now leads with the user's local wall
    clock (e.g. now Thu, Jun 18, 17:53 America/Los_Angeles) so the agent can reason about
    time-of-day, scheduling, and deadlines. Labelled distinctly from quota to avoid the
    clock-time/budget conflation the stamp already guards against.
  • Reset-aware quota guidance. Mid-task advice no longer says "descend / slow down" when the
    5h window will RESET before the velocity engine projects exhaustion (or a reset is ≤10 min
    away) — the real risk is running dry before the reset, not a low %. Deterministic from
    resets_at + projected_exhaustion; optimism requires a positive signal, so unknown burn
    stays cautious.