v0.5.0 — Epic update: collaboration visualization

Epic update: collaboration visualization is here.

The collaboration workbench makes reciprocal, task-bound collaboration between Codex, Claude Code, and Antigravity visible and actionable. Follow the research question, inspect original evidence, and manage assignments from one local interface.

What's new

• A complete collaboration workspace: overview, directed message map, task board and list, evidence records with dependencies, activity, filters, and a detailed inspector.
• Explicit task controls: create scoped assignments, review preflight, dispatch with an explicit model, send messages and replies, and inspect operation results. Creation does not start a provider.
• Pause new work across the shared bridge: hold new executions and automatic follow-up while already admitted calls finish. Messages and results remain available; resume preserves existing routes and sessions.
• Sources and status that keep their meaning: complete briefs and original messages, Unicode-safe source paging, loaded-text export, requested versus reported model identity, missing-source states, and clear distinctions between delivery, validation records, and scientific acceptance.
• A polished English interface: responsive layouts, keyboard navigation, accessible controls, and an offline, clearly labelled fictional demo. No frontend build, external font service, or CDN is required.

This builds on the existing bidirectional information exchange and configured autonomous wakeup paths. It does not claim that a new interface or protocol rule has been shown to improve research quality.

Start or upgrade

Python 3.11+ and Node.js 24+ are required for the workbench. From the complete bundle:

python3 setup.py workbench --demo --open
python3 setup.py install --project ../my-research --upgrade

From an installed research workspace:

python3 .agents/skills/codex-research-orchestrator/scripts/research_setup.py workbench --project . --open

Upgrade retains research records, memory, bridge data, and client configuration. Reopen connected MCP clients after upgrading so all bridge processes use the new pause guard. The server binds only to loopback and uses a per-run private access URL. Pause does not terminate running calls or control unrelated jobs.

Validation

Nine workbench tests, the existing bridge suite, and 29 setup tests passed. The browser walkthrough covered desktop and narrow layouts, source inspection, task creation, messaging during pause, resume, and simulated dispatch. All new validation used isolated synthetic data and no paid model or live provider calls. See validation and the workbench guide.

Packages

• codex-research-orchestrator-v0.5.0.zip: complete repository bundle, recommended.
• codex-research-orchestrator-skill-v0.5.0.zip: standalone research skill and setup/memory helpers; the workbench requires the separate runtime.
• research-agent-bridge-v0.5.0.zip: MCP runtime with the workbench.

Created by Aoudsung. MIT licensed. An independent community project with no OpenAI affiliation or endorsement.

v0.4.0 — Focused authorship and constructive research collaboration

This release updates the research collaboration protocol for Codex while retaining optional reciprocal collaboration across Claude Code, Antigravity, and Codex.

What changes

• Load only the relevant duties for six roles: researcher, optional brainstormer, reviewer, implementer, collaboration optimizer, and fact observer. Project settings bind the actual participants and models.
• When direction is unclear, form short independent constructions before comparison, then assign a main author and complementary evidence questions.
• For original research, require rigorous support and a substantive contribution. Counterexamples lead to a concrete reconstruction within the original object; a ruled-out route is retained as evidence. Reproduction, surveys, bounded reviews, and engineering support keep their own completion criteria.
• Separate fact collection from collaboration diagnosis. Fact-only checks report source-backed changes or UNCHANGED, without ratings or recommendations. Observation reuses existing bindings and baselines and does not start monitoring by default.
• Distinguish handling a review finding, resolving the original difficulty, and establishing a contribution in existing project records. Keep handoffs incremental and independent inputs isolated.

Compatibility and downloads

No Python CLI, memory kind, database field, or MCP protocol changes. Existing projects need no migration or rebuilt records. Use the existing install --upgrade command; research records and configuration are preserved. The optional runtime package is included with synchronized version metadata only.

• Recommended: codex-research-orchestrator-v0.4.0.zip — complete bundle and guided installer.
• codex-research-orchestrator-skill-v0.4.0.zip — standalone research skill and helpers.
• research-agent-bridge-v0.4.0.zip — optional MCP runtime.

Validation

The skill validator and targeted release checks passed, including six role resources and 91 relative references and anchors. Three short author walkthroughs covered an original candidate facing a counterexample, a completed reproduction, and a fact-only observation.

These are instruction and package checks, not new model-backed behavioral trials or evidence of higher research quality. No client communication, model, training, or performance tests were rerun. See the v0.4 validation record.

English release. Created by Aoudsung under the MIT License.

v0.3.1 — Real-client validation and integration fixes

This patch release follows a bounded live-client check of reciprocal communication among Codex, Claude Code, and Antigravity.

All six directed request paths returned matching MCP acknowledgements after event-triggered continuation of the original bound sessions. The test used Codex App Server 0.164.0 with Luna low, Claude Code 2.1.310 with Haiku 5.5 low, and Antigravity CLI 1.4.7 with Gemini 3.8 Flash Low. It used a separate project, database, tasks, and configured custom_command host adapters. All test routes and dedicated receiver processes were stopped afterward.

Fixes

• Apply Antigravity CLI/IDE workspace configuration at .agents/mcp_config.json, with its supported tool timeout field.
• Request Antigravity streaming JSON and parse the current nested model and result envelopes.
• Detect tool permission denials even when the client exits zero or reports a successful terminal status.
• Read unambiguous Claude runtime model evidence from modelUsage.
• Expose explicit reasoning effort for Claude execution/resumption and Antigravity CLI execution.
• Clarify tool grants, including reply permissions, and the distinction between configuration, delivery, and actual processing.

All 27 onboarding regression tests and the affected bridge suite passed. New fixtures protect the real-client output shapes and denied-tool cases.

This verifies real model-backed sessions through explicitly configured host adapters. It does not claim native idle-window activation in the graphical Codex Desktop or Antigravity applications, or automatic wakeup without an adapter. See the live-client validation record for exact scope and evidence.

Download the complete bundle for guided installation. Standalone skill and runtime archives are also included. English documentation; MIT License.

v0.3.0 — Guided research setup and cross-agent collaboration

This release makes the Codex research skill easier to install and use, including its optional MCP bridge for autonomous, reciprocal information exchange and task collaboration among Claude Code, Antigravity, and Codex.

Start with a research goal

Install the complete bundle into a project, open it in Codex, and ask $codex-research-orchestrator to investigate your question. The skill handles minimal records and memory, then continues the next authorized research action. Natural-language entry points cover starting, resuming, progress, connecting peers, and troubleshooting.

What changed

• Added install, init, connect, and read-only doctor commands with clear next actions.
• Installed skill and runtime files stay in the research workspace and work after the download is removed.
• Initialization preserves the original question and creates research memory. Existing research is retained, including during concurrent initialization.
• Connection setup generates actual paths and native project configuration for Codex and Claude Code, with suitable delivery-wait timeouts. Antigravity receives a JSON export for import through its supported interface.
• Upgrades back up managed files and preserve research documents, memory, bridge data, and unrelated client settings.
• Rewrote the English quickstart, first-use guide, connection instructions, and troubleshooting around the installed workspace.

The bridge supports peers exchanging requests, evidence, questions, revisions, and results within the same authorized task. Research roles remain configurable. The lead retains scientific decisions and final acceptance; a message does not expand permissions or prove a scientific claim.

Downloads

• Recommended: codex-research-orchestrator-v0.3.0.zip — complete bundle with the guided installer and optional runtime.
• codex-research-orchestrator-skill-v0.3.0.zip — standalone skill and memory helper.
• research-agent-bridge-v0.3.0.zip — runtime for users managing MCP configuration separately.

Setup requires Python 3.11+. Optional bridge connections require Node.js 24+. Provider execution uses each provider's CLI and normal account setup. Installation does not create accounts, launch models, edit global configuration, or enable background wakeup.

Validation

All 27 onboarding regression tests passed on an isolated Linux server. Six generated client-to-adapter configurations started real MCP stdio processes and completed initialization and tool discovery after the source download was deleted. Existing sample settings were preserved.

These checks did not run paid provider models or verify real desktop login, host loading, or idle-chat wakeup. See v0.3 validation for scope and evidence.

Created by Aoudsung. Released under the MIT License.

v0.2.0 — Autonomous cross-agent collaboration

Autonomous collaboration across Claude Code, Antigravity, and Codex

v0.2 adds an optional shared MCP runtime so Claude Code, Antigravity, and Codex can exchange requests, evidence, revisions, and results and continue authorized work with less manual message relay.

The focus is reciprocal collaboration: any configured participant can initiate a permitted exchange, ask for clarification, return an artifact, or send focused feedback to the same task. Durable bindings and incremental delivery keep the work connected across turns and provider boundaries.

What is included

• Three MCP server adapters with a shared message, task, patch, and validation store.
• Explicit task/session/model bindings, with requested identity separated from runtime-reported evidence.
• Cursor-based messages, compact delivery summaries, structured patches, and revision-linked validation feedback.
• Distinct states for inbox updates, complete artifact delivery, host acceptance, and target processing.
• Updated skill guidance for parallel partners, external authors, independent review, and event-driven continuation.
• English setup documentation and sanitized examples. Private task seeds, personal paths, research state, and account configuration are excluded.

Correctness and verification

The isolated Linux suite passed on Node.js v24.22.1. It exercises three real MCP stdio processes, synthetic provider executables, concurrent event publication, scoped patch paths, execution preflight, delivery manifests, repair feedback, and controlled host-receipt behavior.

Independent review findings were fixed, including event-cursor races, patch path/rename scope checks, preflight bypasses, and unintended route or daemon changes.

These tests do not claim live model execution, autonomous scientific benefit, or a real idle Desktop wakeup for every host. Configure and verify the host route before relying on cross-turn continuation. Full details: v0.2 validation.

Downloads

• Standalone skill: the Codex research workflow and memory helper.
• Agent bridge: the optional MCP runtime, source, synthetic tests, and setup examples.
• Full repository: both components and all documentation.

The runtime requires Node.js 24+ and an explicit shared AGENT_BRIDGE_DB path. Provider execution also requires the selected provider CLI and its normal local authentication. Installing the skill does not register providers, grant permissions, or start background processes.

MIT licensed. Existing research methods, jobs, and the author's live MCP installation were not changed by this release.

v0.1.1 — Codex Research Orchestrator

Context packets now show what they leave out, and handoffs get stricter rules for output formats and state updates.

Changes:

• Every context packet ends with a "Not shown" section. It counts every admitted record left out of the packet and indexes as many as fit in the budget. Resume entries give record IDs and titles; independent entries use neutral labels, so titles and paths stay in the lead manifest.
• New --fill-recent option: after query matches, admitted records without a lexical match fill the remaining budget, newest first.
• The assembly time moved from the first line to the last, so packets rebuilt from the same inputs share all earlier text.
• References add rules for output contracts and repair (full contract and one valid example in the first request; repairs state the violation instead of pasting the rejected output), inputs for each step, whole-state updates, failure history, and incremental work-index entries.
• The memory-tool guarantees now state the conditions they depend on.

Validation:

• 25 memory/context regression tests (the 21 from 0.1.0 plus 4 new) passed on macOS with Python 3.13.1 / SQLite 3.45.3 and with Python 3.10.20 / SQLite 3.53.1.
• In the synthetic 1,000-record check, the packet used 11,975 of its 12,000-character budget; the extra space holds the omission index. The CLI receipt was 203 characters.
• See docs/validation.md for scope and limitations. These checks do not establish comparative research quality or semantic-retrieval performance.

Memory helper requirements are unchanged: Python 3.10+ with SQLite FTS5, with no model API, embedding download, or background service.

Install the standalone archive as a skill, or use skills/codex-research-orchestrator from the full repository.

v0.1.0 — Codex Research Orchestrator

Initial English release of Codex Research Orchestrator.

A research collaboration skill built specifically for Codex, focused on scientific problem selection, continuity of research purpose, and alignment with user requirements.

Included:

• Configurable research roles, independent review, implementation handoffs, and persistent records.
• Local source-backed memory with SQLite FTS5, explicit dependencies, and invalidation of stale conclusions.
• Task-specific context packets with fresh controls, independent-material allowlists, whole-record character budgets, and compact tool responses.
• Design rationale, comparisons with five research automation projects, and adaptations from Mem0, QMD, Letta, Graphiti, and claude-mem.
• MIT license, project templates, a standalone skill archive, and the complete repository/plugin archive.

Validation:

• 21 memory/context regression tests passed on Python 3.10.12 / SQLite 3.37.2.
• In a synthetic 1,000-record check, a 982,704-character source corpus produced an 11,548-character packet under a 12,000-character budget; the CLI receipt was 172 characters.
• See docs/validation.md for exact scope and limitations. These checks do not establish comparative research-quality or semantic-retrieval performance.

Memory helper requirements: Python 3.10+ with SQLite FTS5. No model API, embedding download, or background service is required by the helper.

Install the standalone archive as a skill, or use skills/codex-research-orchestrator from the full repository.