nyldn/claude-octopus v11.16.0

on GitHub

Fixed

• Decision support now includes the cost of doing nothing, sets criteria and
  weights before scoring, and shows raw and weighted scores. Presentation
  templates no longer recommend the first option before analysis. Uncertain
  decisions can include an optional scenario stress test.
  Fixes #1220.
  #1247.

• Workflow skills can now persist phase state through update_state, query
  completed phases and import definition synthesis into the roadmap. History
  notes cannot change a phase's recorded status. Initializing a project with
  a missing state file preserves existing issues, lessons and phase plans.
  Fixes #1169.
  #1248.

• Configured fallback chains now honor OCTO_ALLOWED_PROVIDERS and the
  session or global provider allowlist. run_agent_sync_fallback_chain
  dispatched its primary, preferred fallback and chain candidates without
  checking the allowlist. With OCTO_ALLOWED_PROVIDERS="codex claude", Tangle
  decomposition still sent the prompt to its default agy researcher seat
  first. The chain now skips any candidate outside the allowlist and logs the
  skip. If no candidate is allowed, it fails closed without dispatching.
  octo_fallback_agent_available applies the same check, so
  get_fallback_agent and octo_fallback_first_available no longer return a
  disallowed provider just because its CLI is installed.
  #1244.

• Preflight checks the council roster that OCTOPUS_COUNCIL_DEFAULT_PROVIDERS
  pins. A Codex-led session pins claude,agy, but preflight ignored the list:
  the static pass checked every non-Claude provider, the smoke test always
  tried Codex, and nothing reported on the Claude seat. With a roster set,
  orchestrate.sh preflight checks exactly those seats, Claude included, names
  the list it used, and fails naming any seat that is not ready. The smoke test
  and /octo:preflight cover only the roster, and the smoke test fails when
  any roster seat it tested fails. A pinned Codex seat with an expired OAuth
  token fails the roster even when other seats are ready. A seat with no
  provider detection, such as openai-tools, is listed as not checked instead
  of failing the roster or counting as degraded. The JSON gains roster, roster_unchecked and sandbox, and cached results are
  keyed on the roster. Unset or auto keeps the previous behavior.

• Preflight readiness inside a macOS sandbox. The /octo:preflight loop sent
  stderr through 2> >(cat >&2), a /dev/fd write that Codex and Claude Code
  sandboxes deny, so every provider read check-failed there. Inside such a
  sandbox (CODEX_SANDBOX or SANDBOX_RUNTIME), a provider that is not ready
  now carries a note that keychain logins, network access or files outside the
  workspace can fail there, with the remedy to re-run preflight outside it.
  #1229.

Maintenance

• Updated the MCP server's Zod dependency.
  #1235.
• The timer regression waits for completion before checking failure precedence.
  #1249.