nyldn/claude-octopus v11.6.0

v11.6.0 — Smarter task planning, safer context routing, and more accurate diagnostics

on GitHub

[11.6.0] - 2026-09-18

Added

• Tangle can adapt its write scope to the task and explicitly authorize
  external read context, making parallel planning safer and more useful
  (#1043, #1044).

Changed

• Context summarization now uses the configured feature seats and preserves the
  target workflow's context budget during preflight (#1042, #1045).
• Structured Tangle decomposition is normalized before the compatibility
  fallback path, improving planning reliability across provider output formats
  (#1046).

Fixed

• Doctor now finds the installed plugin and configured workspace when invoked
  directly or through the early orchestrate.sh doctor dispatch, so healthy
  installations no longer report path-related failures (#1048, #1049).

• The Tangle quality gate uses the correction-overlay rate so review results
  reflect the actual corrected output (#1038).

• Council verdict extraction (_council_parse_final_verdict) no longer misreads
  a seat's VERDICT: APPROVE as REVISE, which was systematically corrupting the
  quorum tally (met: false, distinct_approving_providers: 0) even when the raw
  seat bodies clearly approved (#2346). The runner wraps each seat's output in a
  provenance envelope (## UNVERIFIED CONSULTATIVE OUTPUT header +
  <external-cli-output>…</external-cli-output> + ## END … footer), so the
  verdict is no longer the file's final line — and the parser treated the closing
  wrapper lines as trailing review content, demoting the final verdict to nothing.
  The parser now skips those runner-added envelope wrapper lines so the verdict
  inside is recognized as final, while deliberately keeping the strict contract
  intact: the verdict must still be the last top-level line, the sole declaration,
  and end-anchored, so a quoted/fenced example verdict or an unfinished
  (trailing-content) verdict still cannot vote. A degenerate/missing chair already
  never forced met: false; a regression test locks in that a mixed-vendor pair of
  enveloped VERDICT: APPROVE seats yields distinct_approving_providers: 2 /
  met: true alongside a degenerate chair.

• Remove the matcher-group if keys from the freeze-check and telemetry-webhook entries in hooks/hooks.json. Claude Code only reads if on individual hook handlers, so it ignored both keys and printed hooks.json: unknown keys "if" in hooks.PreToolUse[6], "if" in hooks.PostToolUse[3] ignored at startup. Both scripts already check freeze state and the webhook URL themselves, so hook behavior is unchanged.