v1.8.1

What's Changed

Bug Fixes

• Headless idle-detection completion state: Runs that finished via idle detection were incorrectly marked as errors in metadata.json, poisoning amicus_list, amicus_status, and wave rollups. Both idle-completion exits (SDK-authoritative session.status idle signal and stable-poll activity heuristic) now properly set the completed flag, matching the fold-marker behavior.
• Abort precedence in terminal state resolution: Clarified that abort-over-completed takes precedence in resolveTerminalState.

Documentation & Skills

• Plugin quick-start accuracy: Clarified that plugin installs do not put amicus on PATH; all CLI calls must use npx -y amicus@latest <command>. Both skills now include an npx-fallback transport rule.
• Command and flag documentation: README and docs/usage.md now document doctor, key, and council subcommands with full syntax. Troubleshooting leads with amicus doctor instead of stale port-checking guidance; replaced false "amicus list shows active servers" with real netstat/lsof commands.
• Council mechanics hardening: Stage-2/3 judge briefings now require a mandatory no-tools preamble (closes tool-wandering and anonymization leaks). Added scratch-cwd advice for defense-in-depth. Budget-gate pass-through documented for repair and chair calls (--max-cost/--no-cost-gate must travel with the council wave, repairs, and chair).
• Tally schema completeness: Stage-2 assembly recipe now leads with step 0 (build meta and findings[] first) and includes a five-keys checklist to prevent BAD_ARGS failures. Clarified that the raiser's own adjudication is included in the input.
• Stage-5 artifact defaults: report.html is now the mandatory default final council artifact (self-contained, shareable page). Stage 5 now requires an inline verdict summary in chat (chair's assessment, tier counts, what was applied), never file paths alone.
• Second-opinion skill scope: Removed "second opinion from another model" from the sidecar frontmatter; those requests now route to the second-opinion skill instead, clarifying the role boundary.
• MODEL-NOTES field lessons: Absorbed durable machine-independent lessons from council runs 4-7: PowerShell --models quoting rule (now in every example); current-date injection for time-sensitive artifacts; long-agentic-read model selection guidance; new per-model sections for Grok, Kimi, Mistral, and Claude-in-council with reliability data and caveats.
• Sidecar SKILL.md frontmatter overhaul: Rewrote operating rules for clarity, fixed command/flag accuracy, and confirmed both skills include the npx fallback rule.
• Release checklist added: New guidance in docs/publishing.md for MODEL-NOTES fold-back, version sync between package.json and plugin.json, and the canonical single-commit release ritual.

Tests

• Added comprehensive coverage for docs accuracy: command/flag tables (docs-command-coverage.test.js), plugin quick-start rules (plugin-quickstart-docs.test.js), skill cost-gate pass-through (skill-cost-gate-docs.test.js), council hardening (skill-council-hardening-docs.test.js), judge no-tools preamble (skill-judge-hardening-docs.test.js), sidecar scope (skill-sidecar-docs.test.js), tally five-keys recipe (skill-tally-recipe-docs.test.js), and MODEL-NOTES shipped sections (model-notes-shipped-docs.test.js).
• Headless idle-completion test: New headless-idle-completion.test.js pins both idle-detection exits (SDK signal and stable-poll heuristic) as completed, not error.

Full Changelog: v1.8.0...v1.8.1