Portable LLM Council v0.2.6

Portable LLM Council v0.2.6

Improve advisor depth, review accuracy, conditional recommendations, and recovery after material input defects.

• Give advisors a 200–400 word target, with up to 500 for consequential detail; retain concise reviews.
• Review different contributions fairly and check factual criticisms and claims of missing coverage.
• Carry decisive conditions into the chair’s opening verdict and qualify unresolved feasibility.
• Strengthen packet assembly checks and bounded recovery for material input defects across all roles.
• Check decision-critical arithmetic, dates, units, and agreement before delivery while preserving original outputs.
• Keep native role definitions and historical website examples unchanged.

The five-advisor, five-reviewer, separate-chair method is unchanged. All seven harnesses receive the updated instructions through the shared assignments; their 21 native role definitions need no changes. No new runtime council or model-performance comparison is claimed. The food-rescue run and website scenario selection are deferred to a separate publication review.

Portable LLM Council v0.2.5

Portable LLM Council v0.2.5

Clarify the council’s business and technical purpose and its OpenAI installation scope.

• Lead website copy and sharing imagery with pricing, hiring, product, operations, and technology decisions.
• Use “AI assistant” in reader-facing installation guidance.
• Explain Codex workflows in the ChatGPT desktop app, CLI, and IDE extension while keeping the Codex adapter and installation paths.
• Describe ChatGPT Work subagents as documented capability awaiting package-specific installation and workflow validation.

Council methodology and native agent definitions are unchanged. The Codex adapter documentation is updated; no new ChatGPT Work runtime compatibility is claimed.

Portable LLM Council v0.2.4

Portable LLM Council v0.2.4

This focused follow-up makes dispatch records more compact and honest about their provenance without shortening worker inputs.

• Allow shared record components to be stored once; distinguish verified capture, coordinator copies, and reconstructions.
• Preserve reviewer presentation order across retries, identify the completed attempt used, and allow unexposed worker IDs.
• Clarify the historical report link and require phone-friendly future reports.
• Improve mobile navigation targets, remove redundant Droid wording, link directly to dispatch templates, and clarify Claude git-status context.

The five-advisor, five-reviewer, separate-chair method and historical example records remain unchanged. Build, source consistency, links, clipboard behavior, and all three website pages at desktop, 390px, and 320px were checked. No new council or Claude runtime test was run.

Portable LLM Council v0.2.3

Portable LLM Council v0.2.3

Council runs now use explicit dispatch templates and preserve the exact inputs sent to each worker. Reviewers receive complete, minimally anonymized answers in rotated order while answer identities remain fixed.

• Reinforce supplied-input boundaries across all 21 native role definitions.
• Request Claude Code instruction omission (v2.1.271+) and Codex read-only execution, with effective permissions and isolation still subject to runtime verification.
• Clarify adapter execution notes, shared-file exposure, and UTC artifact filenames.
• Improve the website introduction, navigation, installation journey, and sharing previews. Link the original example report unchanged and retain its limitations.

The five-advisor, five-reviewer, separate-chair structure is unchanged. The preserved example predates these refinements. Build, static checks, desktop/mobile layouts, links, and clipboard behavior were checked; the updated native definitions have not undergone a new cross-harness council run.

Portable LLM Council v0.2.2 — workflow refinements

Portable LLM Council v0.2.2 — workflow refinements

This release refines when the council runs and how it handles missing context, interrupted work, and incomplete results.

• Narrow the trigger so quick critiques do not automatically start a full council.
• Ask for decisive missing information only when available context or an explicit conditional assumption cannot resolve it.
• Recover outstanding worker results through the supported lifecycle before declaring a round incomplete.
• Preserve returned work in both HTML and Markdown artifacts when recovery is impossible, clearly marked incomplete without inventing a verdict.
• Reuse capability checks within an unchanged session while retaining reviewer-input checks and disclosure of unverified settings.
• Clarify adapter-specific fallback boundaries and normal host approval handling without permitting policy bypasses.

The method remains five advisors, five fresh blind reviewers, and a separate chair. Both output formats remain required. Native agent definitions, model inheritance, and supported harness targets are unchanged.

Updating: Refresh the shared llm-council skill directory, including its references. Native agent files do not need replacement for this release.

Validation: Staging and repository copies match; local links, YAML frontmatter, and whitespace checks passed. No new live council or cross-harness runtime tests were performed for this update.

Portable LLM Council v0.2.1 — assistant-led installation guidance

Portable LLM Council v0.2.1 — assistant-led installation guidance

The README now includes a copyable prompt asking your assistant to install the shared skill and the native agents for its harness. It covers backing up existing copies and verifying discovery, with guidance on approval prompts and starting a new session.

Manual installation instructions remain available. This is a documentation-only update; existing users do not need to reinstall the skill or agents.

Portable LLM Council v0.2.0 — Claude support

Portable LLM Council v0.2.0 — Claude support

Adds Claude Code support, including local Code sessions in Claude Desktop. The package now provides seven harness adapters and 21 native role definitions around one shared council skill.

• Three Claude roles: advisor, reviewer, and chair, used for eleven separate worker runs.
• Fresh non-fork dispatch guidance, text-only worker configuration, model inheritance, and explicit isolation limits.
• Personal copy/symlink options and project installation instructions.
• Installation and compatibility guidance; existing council assignments and the other eighteen role files are unchanged.

Claude support covers Claude Code and local Code sessions in Claude Desktop. Cowork, ordinary Chat, and cloud deployment are outside this adapter's scope.

Install or update

Follow the repository README for your harness. Existing users should back up and deliberately replace the shared skill to receive the Claude reference; fresh-install commands skip existing files. Install only the Claude roles if adding Claude support. No other harness's role files changed.

GitHub supplies the source ZIP and tar.gz automatically. Update existing installations explicitly using the README instructions.

Portable LLM Council v0.1.0

Portable LLM Council v0.1.0

A shared council skill with native subagent definitions for Codex, Cursor, Command Code, Grok Build, OpenCode, and Factory Droid. Adapted from Ole Lehmann's Claude-focused lens-based workflow, drawing on Andrej Karpathy's LLM Council.

Five advisors, five fresh blind reviewers, and one chair produce an HTML report and a Markdown transcript. No specific model is pinned. See the README for installation and the validation notes for documented capabilities and untested behavior.

This initial release has no comprehensive model or end-to-end harness test matrix. Claude Code and ordinary chat products are not included targets.

Download GitHub’s automatic “Source code (zip)” asset below and follow its README, or clone the repository.