nyldn/claude-octopus v10.0.0

v10.0.0 — Reliability modernization

on GitHub

Added

• A schema-versioned execution contract records every synchronous, supervised,
  and Agent Teams seat from planned through a terminal state. New
  octopus status --run and octopus explain --run commands reconstruct runs
  from durable artifacts without calling a provider.
• Setup and Doctor 2.0 now distinguish installation, authentication, model
  readiness, plugin assembly, cache/state writability, stale runs, and orphan
  process evidence. doctor --json emits a stable v10 machine-readable report.
• Provider Registry 2.0 makes canonical identity, aliases, authentication mode,
  health and detection handlers, model environment, context budget, cost class,
  sandbox class, and independence organization parity-enforced data.
• Safe cancellation and recovery track process groups, verify descendant
  cleanup, reconcile stale running seats, retain source/worktree attribution,
  and allow retries only as distinct attempts for non-contributed work.
• Eval-backed routing classifies mechanical, balanced, premium, review, and
  security work while preserving user and project pins. Fable escalation now
  has an atomic one-seat-per-run claim, a 524,288-byte default input ceiling,
  and cross-vendor verification rules.
• A hermetic end-to-end failure-injection suite drives the real orchestration
  entrypoint through success, auth and health rejection, unusable output,
  timeout and partial timeout, child-process cleanup, refusal, signal retry,
  closed stdin, and persistence failure with zero provider billing.

Changed

• A provider response contributes to synthesis only after its output is durable
  and validated. Empty, placeholder, cancelled, timed-out, and unpersisted
  results remain visible but ineligible.
• Cache success messages now reflect actual persistence; a cache write failure
  preserves usable synthesis output and records a warning instead of claiming
  the result was cached.
• Doctor returns 1 for diagnostic failures and 2 for invalid arguments.
  JSON stdout remains valid for diagnostic failures, so scripts should capture
  the process status separately from the report body.

Fixed

• Synchronous Antigravity seats now retry once after an intermittent exit 139
  (SIGSEGV) without extending the original timeout budget. Signal-terminated
  seats retain private stderr artifacts linked from run status, including when
  the retry succeeds, so orchestration failures remain diagnosable.