Codex Council v0.7.0 - Bounded Role Tuning

Highlights

• Added codex-council-alters, a guided skill for tuning council member behavior.
• Added bounded local role tuning for Ada, Grace, Hypatia, Florence, Turing, Seymour, and Leonardo.
• Bob is explicitly excluded because he is an evidence runner, not a council member or reviewer.
• Added CLI support:
  • alters list
  • alters show
  • alters preview
  • alters configure
  • alters reset
• Role tuning is injected into future council prompts as a compact advisory block.
• Added guardrails so tuning cannot override blockers, dissent, verification, anonymization, preflight, safety checks, or role boundaries.
• Added rejection for unsafe tuning such as “always approve”, “hide blockers”, or “skip verification”.
• Added token accounting for alter_tuning_prompt_tokens in preflight estimates.
• Added stable state handling for versioned plugin-cache installs, so role tuning and profile history survive updates.
• Added automatic migration from old version-local state into the stable plugin state directory.
• Added council lifecycle guidance: after the six members finish, collect outputs and close them before spawning reviewers, Leonardo, Bob, or any extra agents.

Safety

• Custom role tuning is advisory only.
• Raw tuning instructions are not written to invocation logs.
• Session metadata stores compact tuning summaries/fingerprints, not private raw instructions.
• Corrupt alter config blocks execution instead of silently resetting behavior.

Verification

• python3 -m unittest discover -s tests -v
• python3 scripts/codex_council.py validate --plugin-root . --strict
• git diff --check

Update

npx codex-marketplace add ercoledevs/codex-council --plugin --global -y

Codex Council v0.6.0 - Safer token accounting and typed synthesis

Highlights

• Added safer pre/post token accounting with comparable pre_execution_estimate, post_execution_estimate, and artifact_only_tokens.
• Added persistent preflight reports: preflight-estimate.json and preflight-estimate.md.
• Added prompt logging for council members, reviewers, and chairman synthesis to improve post-session estimation.
• Moved council runtime state to plugin-local .codex-council/ storage instead of writing temporary data into project repositories.
• Added typed council sessions with --type architecture|implementation|decision|skill|frontend.
• Added compact --skill-review mode with focused skill/tool review lenses.
• Added a separate chairman synthesis pass with explicit synthesis inputs.
• Added invocation classification to avoid accidentally launching the council for meta/reference questions.
• Added sanitized invocation logging and optional path-only raw bundles.

Token Safety

• Post-session stats no longer treat saved artifacts as the main token cost.
• Reports now separate session-wide estimates from artifact-only tokens.
• Missing prompt/output data is reported as partial coverage instead of being presented as complete billing data.
• No billing/API usage is claimed unless it is actually measurable.

Verification

• python3 -m unittest discover -s tests -v
• python3 scripts/codex_council.py validate --plugin-root . --strict
• git diff --check

v0.5.0 - Council Banner and Session Stats

What's Changed

This release adds visible session polish and useful closeout reporting for Codex Council.

Added

• Added a compact ASCII Council banner for session starts.
• Added stats CLI command for end-of-session reporting.
• Added stats --json for machine-readable output.
• Added stats --write to persist stats.json and stats.md.
• Added estimated artifact-token reporting based on local session files.
• Added session-level counts for:
  • roles
  • reviewers
  • evidence runners
  • artifact files
  • member files
  • review files
  • evidence files
• Added largest-artifact breakdown to help spot noisy session files.
• Added chat-specific guidance: when Codex runs the Council in chat, the banner and stats should be relayed in the conversation, not hidden in shell output.

Improved

• Council sessions now distinguish between terminal-visible and chat-visible banners.
• Stats now clearly state that token counts are estimates from local artifacts only.
• The skill contract now requires persisted compact artifacts before claiming a full session stats report.
• If member or reviewer output only exists in chat, stats must be treated as scaffold-only.
• README and reference docs now explain how to use session stats safely.

Important Note

Estimated artifact tokens are not actual Codex token usage, billing telemetry, hidden prompt overhead, or tool-call accounting. They are local estimates calculated from saved session files.

Validation

Verified before release with:

python3 -m unittest discover -s tests -v
python3 scripts/codex_council.py validate --plugin-root . --strict
python3 scripts/codex_council.py check-update --latest-version v0.5.0 --json
git diff --check

Result:

• 24 tests passing
• strict plugin validation passing
• update-check path passing
• diff hygiene clean
• no private paths or secrets detected in the release scan

Upgrade

npx codex-marketplace add ercoledevs/codex-council --plugin --global -y

For project-local installs:

npx codex-marketplace add ercoledevs/codex-council --plugin --project -y

v0.4.0 - Performance Council Member

What's Changed

Codex Council now includes a dedicated performance perspective in the core council flow.

This release adds a new council member focused on technical performance impact, expands the reviewer layer, and improves the final synthesis so architectural, implementation, UX, security, maintainability, and performance concerns are considered together before the main agent receives a recommendation.

Added

• Added Seymour Cray - Performance Engineer as a core council member.
• Expanded proposal scoring from A-E to A-F to include the new performance perspective.
• Added performance review coverage for:
  • latency
  • throughput
  • memory usage
  • CPU cost
  • I/O cost
  • database/query impact
  • cache behavior
  • startup/build time
  • concurrency
  • scalability
  • operational cost
• Added Performance Impact and Measurement Required sections to council outputs.
• Added a new performance-impact-reviewer.
• Added a new coverage-integrator reviewer.
• Expanded deep review mode so more reviewer perspectives are checked before final synthesis.
• Added performance-sensitive workflow guidance and competency-pack documentation.

Improved

• The final council synthesis now has stronger cross-role coverage.
• Performance risks are separated from general engineering concerns, making tradeoffs easier to evaluate.
• Reviewer output is better structured so the main agent receives a more complete, balanced recommendation.
• Token efficiency rules were preserved while adding the new member, keeping the skill compact and focused.

Validation

Verified before release with:

python3 -m unittest discover -s tests -v
python3 scripts/codex_council.py validate --plugin-root /Users/s.ercole/plugins/codex-council --strict
python3 scripts/codex_council.py check-update --latest-version v0.4.1 --json
git diff --check

Result:

• 20 tests passing
• strict plugin validation passing
• update-check path passing
• leak scan clean
• diff hygiene clean

Upgrade

To update the plugin:

npx codex-marketplace add ercoledevs/codex-council --plugin --global -y

Or install it inside a specific project if you prefer project-local usage.

v0.3.0 - Frontend UX gate, Bob browser evidence, and token-efficient council flow

Codex Council v0.3.0

This release improves Codex Council with a new frontend/UX review path and a more token-efficient council flow.

The goal is simple: keep the same quality gates, blockers, dissent, verification, confidence, and anonymous review, while loading less context and producing less unnecessary text.

Highlights

• Added optional frontend/UX review with Leonardo da Vinci
• Added browser evidence runner Bob
• Added token profiles: compact, balanced, expanded
• Reduced SKILL.md from about 995 words to about 475 words
• Moved detailed guidance into lazy-loaded references
• Strengthened strict validation so token savings do not remove quality gates
• Improved scoring safety by rejecting incomplete candidate coverage

Added

Frontend/UX Review Gate

Frontend reviews can now be created with:

python3 scripts/codex_council.py init \
  --topic "Frontend Modal Review" \
  --root . \
  --mode standard \
  --token-budget compact \
  --frontend-review

This adds two optional frontend-specific roles:

• Leonardo da Vinci - Brutally Honest UX/UI Critic

  • reviews UI/UX decisions
  • catches counterintuitive interactions, weak hierarchy, accessibility gaps, mobile issues, and confusing flows
  • acts as a reviewer/gate, not as a sixth core council member

• Bob - Browser Customer Tester

  • runs browser-oriented evidence checks
  • verifies concrete UI cases suggested by the council
  • reports pass/fail/not-verified evidence
  • does not vote, score, or become a council candidate

Token Profiles

New token budget profiles:

• compact: default profile for normal reviews
• balanced: for ambiguity, meaningful blockers, or non-trivial tradeoffs
• expanded: for security, migrations, data loss, irreversible changes, or audit-heavy work

Example:

python3 scripts/codex_council.py init \
  --topic "Architecture Review" \
  --root . \
  --mode standard \
  --token-budget compact

New Reference Files

Added lazy-loaded references:

• execution-protocol.md
• frontend-ux-browser.md

These keep the main skill smaller while preserving detailed guidance when needed.

Changed

Smaller Main Skill

SKILL.md is now a compact routing kernel.

Detailed stage instructions, frontend evidence rules, scoring details, and token tactics now live in reference files and are loaded only when relevant.

This reduces default context usage while keeping the full workflow available.

Stable Historical Personas

The five core council members now use stable historical personas:

• Ada Lovelace - Principal Architect
• Grace Hopper - Reliability Engineer
• Hypatia - Security and Governance Reviewer
• Florence Nightingale - Product and Operator Advocate
• Alan Turing - Contrarian Red Team

This makes council outputs easier to follow even when the UI assigns random agent nicknames.

Frontend Review Is a Flag

Frontend review is implemented as a flag, not a separate mode:

--frontend-review

It can be combined with existing modes such as standard or deep.

Improved

Safer Compact Reviews

Reviewer aggregation now rejects incomplete candidate scoring.

If a reviewer skips an expected candidate, the aggregation fails instead of silently producing a weak or misleading winner.

Frontend Evidence Contract

Frontend sessions now include:

## Frontend Evidence

This separates:

• council judgment
• Leonardo UX critique
• Bob browser evidence

Stronger Validation

validate --strict now checks that:

• SKILL.md stays compact
• heavy schemas stay in references
• blockers are preserved
• dissent is preserved
• verification is preserved
• anonymization is preserved
• missing candidate scores do not pass as valid coverage
• Bob is never treated as a voting council member
• UI behavior is not claimed as verified without browser evidence

Token Optimization

This release applies several token-saving tactics:

• keep SKILL.md short and focused
• lazy-load detailed references only when needed
• use compact as the default profile
• reduce output tokens before removing important evidence
• keep stable instructions before dynamic project context
• escalate to balanced or expanded only when risk requires it

The optimization is designed to reduce token usage without removing the council’s core guarantees.

Verification

Verified with:

python3 -m unittest discover -s tests -v
python3 scripts/codex_council.py validate --plugin-root . --strict
python3 scripts/codex_council.py check-update --latest-version v0.3.1 --json
git diff --check

Results:

• 17 tests passing
• strict plugin validation passing
• update-check JSON passing
• frontend session validation passing
• leak scan clean
• diff hygiene clean

Update

Global install:

npx codex-marketplace add ercoledevs/codex-council --plugin --global -y

Project install:

npx codex-marketplace add ercoledevs/codex-council --plugin --project -y

Then restart or reload Codex.

v0.2.0 - Update checks, release guidance, and Karpathy LLM Council alignment

Highlights

This release improves Codex Council as a public, installable Codex plugin.

Added

• Added check-update CLI command to check GitHub Releases for newer versions.
• Added machine-readable update output with --json.
• Added README guidance for updating the plugin through codex-marketplace.
• Added instructions for users who want release notifications through GitHub Watch -> Custom -> Releases.
• Added a GitHub Release badge to the README.
• Added tests for update-check behavior without requiring live network access.

Improved

• Updated the plugin provenance and documentation to reference:
  • https://github.com/karpathy/llm-council
  • https://llm-council.dev/
• Refined the council workflow around:
  • independent first opinions
  • anonymous peer review/ranking
  • rubric scoring
  • Chairman synthesis
  • explicit dissent and verification handling

Changed

• Bumped plugin version from 0.1.0 to 0.2.0.

Update

If you installed Codex Council globally:

npx codex-marketplace add ercoledevs/codex-council --plugin --global -y