nyldn/claude-octopus v9.51.0

v9.51.0 — Fable 5 dispatch profile + auto-enforced guards

on GitHub

Added

• RELEASING.md: ordered release checklist covering every version-string location, the derived-artifact generators, CI-parity validation, exec-bit checks, fork-PR run approval, the tag-on-merge-commit rule, and GitHub Release creation. Encodes the three CI rounds the v9.50.0 release burned on undocumented generators.
• docs/PROVIDERS.md: provider wiring map. Seven wiring points across five files per provider, with anchors and the traps that have bitten real PRs (case-glob ordering, the two provider-routing whitelists, exec bits, the stdin shim contract, secret-scanner quoting, nested-session markers).
• make sync / make sync-check / make ci-local: one target to regenerate all derived artifacts, one to verify them, and a CI-parity target that mirrors the required checks plus CI-only verifications so local green predicts remote green.
• Executable-bit lint in CI (Portability Lint job): PRs fail if tracked files change mode vs the base branch; the allow-mode-change PR label bypasses intentional changes. Root cause class of PR #579's "Permission denied" failures, now caught pre-merge.
• "Repo Orientation for Agents" section in CLAUDE.md (mirrored in AGENTS.md): derived-artifacts table, hard rules distilled from real CI failures, and a memory ruling for the beads blocked-writes failure mode (do not migrate; record in handoff and flag).
• Fable 5 dispatch profile (skills/blocks/fable5-prompting.md): prompting rules for claude-fable-5 pins, distilled from Anthropic's Fable 5 prompting guide and the fable5-optimizer project. Covers prompt anti-patterns (reasoning-echo asks trigger the reasoning_extraction refusal; token countdowns; aggressive MUST/CRITICAL emphasis; micromanaged step plans), high-effort discipline (the Opus 4.8 xhigh phase table does not carry over), refusal fallback to Opus 4.8, and judgment-vs-mechanical seat routing with a risk-surface escalation list. Wired into CLAUDE.md (cost + effort sections), skill-meta-prompt (model-specific prompt adjustments), octopus-security-audit (never dispatch security passes to Fable 5 — its safety classifiers can refuse adversarial phrasing), and model-cost-compare (risk-surface escalation step, Fable 5 security guardrail).
• Fable 5 mode auto-enforcement (scripts/lib/fable5.sh): detecting a claude-fable-5 pin (OCTOPUS_OPUS_MODEL or OCTOPUS_CLAUDE_SDK_MODEL) now auto-enables three guards with a one-line banner. (1) Security reroute: the model resolver and claude-opus dispatch swap claude-opus-4.8 in for Fable 5 on security dispatches (security-auditor role, squeeze workflow) — its safety classifiers can refuse adversarial security phrasing. (2) Effort clamp: get_effort_level clamps xhigh/max to high for opus-seat Fable pins, including explicit OCTOPUS_EFFORT_OVERRIDE values (Fable 5 effort applies per tool call; higher settings widen scope at 2x cost without extending runs). (3) Refusal retry: claude-sdk-exec.sh retries a refused/empty Fable 5 dispatch once on claude-opus-4-8 (OCTOPUS_FABLE5_NO_RETRY=1 opts out), mirroring the agy silent-empty replay. Master switch OCTOPUS_FABLE5_MODE=auto|off|on. A new SessionStart hook (hooks/fable5-inject.sh) injects the dispatch profile summary when a pin is detected.

Fixed

• OCTOPUS_OPUS_MODEL=claude-fable-5 now reaches the dispatched model flag. The claude-opus dispatch case always passed the bare --model opus alias, which the host resolves to its default Opus — so a Fable 5 pin changed cost labels and resolver output but never the model actually dispatched (the claude-sdk seat was the only real Fable 5 path). The dispatch command now emits --model claude-fable-5 for pinned non-security dispatches and --model claude-opus-4-8 for security dispatches.