nyldn/claude-octopus v11.20.0

v11.20.0 — Faster workflows, clearer presentation outlines and optional web research

on GitHub

Added

• /octo:deck --red-team can review a draft outline through stakeholder seats
  before approval, map decision-blocking objections to the main deck, and prepare
  draft answers and appendix backup slides for Q&A. The pass is opt-in and uses
  the existing debate provider and cost controls. Transports with unbounded
  follow-up requests are excluded from the pass.

• Deck outlines now start with a confirmed governing question, a brief
  situation/complication/question opener and an answer supported by two to four
  pillars. Slide titles form a sentence skeleton that must carry the argument
  in a titles-only read before the user approves rendering. Informational decks
  retain their own purpose. Addresses #1215.

• Executive and board decision decks put the recommendation, context, supporting
  pillars and explicit ask on slide 2. Their outline compares doing nothing
  with a real alternative, ties data-slide highlights to one proof point,
  and includes an appendix for expected questions. Missing decision material
  is identified before outline approval rather than invented. Informational
  updates keep their own purpose, and short decks can combine evidence and
  decision slides within the approved target. The action closes the main deck,
  before appendix backup slides. Addresses #1216.

• Research can optionally share one Keenable web search across its perspectives.
  Set OCTOPUS_RESEARCH_WEB_SEARCH=keenable and include keenable in any active
  provider allowlist. Search suggestions remain untrusted, and lookup failures
  keep the original research prompts. No API key is required for the public
  endpoint. Addresses #1261.

Improved

• Leading CLI help skips workflow initialization. Complete, valid reviews with
  no findings skip verifier and synthesis calls after confirming the diff is unchanged.
• Council advice, critique and revision use bounded concurrency with ordered
  usage reservations. Refinement can overlap its two pinned, independent judges
  under one call budget and deadline.
• Preparation can check an explicit provider panel and validate a short-lived
  readiness receipt. Fresh dispatch authorization remains required.
• Discovery and refinement load phase instructions as needed. Brainstorm, PRD
  and security workflows reuse supplied answers and ask about material gaps.
• Correction loops detect completed workers sooner. Provider retries distinguish
  terminal authentication failures from recoverable refresh or transport failures.
• Worktree metadata sampling, usage reports, context lookup and convergence
  checks avoid repeated subprocesses, whole-log buffering or unchanged file reads.

Fixed

• Worktree progress fingerprints preserve stat compatibility for timestamps before
  1970, including fractional seconds. The batch sampler and shell fallback produce
  the same digest while ordinary files keep the batch sampling path.

• Code review skips complete file diffs using REVIEW.md path globs. Generated
  hunks no longer appear under another file's header, and retained code that
  mentions a skipped path stays intact. Fully skipped diffs launch no reviewers.
  Standard and no-prefix Git diffs are supported, including quoted, binary
  and mode-only paths.
  Fixes #1279.

• Provider smoke tests explain environment construction failures instead of
  failing silently. Search-credential rejection uses the routing error logger,
  with a stderr fallback for standalone callers.

• Optional search skips interpreters without total-deadline support and warns
  when a lookup fails or returns no suggestions, while research continues.

• Codex subscription probes exclude search credentials and skip execution if
  provider environment construction fails.

Full changelog: v11.19.0...v11.20.0