go-agent/codex/IMPROVEMENTS.md
Snider 61e01bfdf1 feat: initial go-agent — agentci + jobrunner + plugins marketplace
Consolidates three codebases into a single agent orchestration repo:

- agentci (from go-scm): Clotho dual-run verification, agent config,
  SSH security (sanitisation, secure commands, token masking)
- jobrunner (from go-scm): Poll-dispatch-report pipeline with 7 handlers
  (dispatch, completion, auto-merge, publish draft, dismiss reviews,
  send fix command, tick parent epic)
- plugins marketplace (from agentic/plugins): 27 Claude/Codex/Gemini
  plugins with shared MCP server

All 150+ tests passing across 6 packages.

Co-Authored-By: Virgil <virgil@lethean.io>
2026-02-21 15:47:19 +00:00

1.9 KiB
Raw Blame History

Codex Extension Improvements (Beyond Claude Capabilities)

Goal

Identify enhancements for the Codex plugin suite that go beyond Claudes current capabilities, while preserving the Axioms of Life ethics modal and the blue-team posture.

Proposed Improvements

  1. MCP-First Commands

    • Replace any shell-bound prompts with MCP tools for safe, policycompliant execution.
    • Provide structured outputs for machinereadable pipelines (JSON summaries, status blocks).
  2. Ethics Modal Enforcement

    • Add a lint check that fails if prompts/tools omit ethics modal references.
    • Provide a codex_ethics_check MCP tool to verify the modal is embedded in outputs.
  3. Strings Safety Scanner

    • Add a guardrail script or MCP tool to flag unsafe string interpolation patterns in diffs.
    • Provide a “safe string” checklist to be autoinserted in risky tasks.
  4. CrossRepo Context Index

    • Build a lightweight index of core-agent plugin commands, scripts, and hooks.
    • Expose a MCP tool codex_index_search to query plugin capabilities.
  5. Deterministic QA Runner

    • Provide MCP tools that wrap core CLI for Go/PHP QA with standardised output.
    • Emit structured results suitable for CI dashboards.
  6. PolicyAware Execution Modes

    • Add command variants that default to “dryrun” and require explicit confirmation.
    • Provide a codex_confirm mechanism for highimpact changes.
  7. Unified Release Metadata

    • Autogenerate a Codex release manifest containing versions, commands, and hashes.
    • Add a “diff since last release” report.
  8. Learning Loop (NonSensitive)

    • Add a mechanism to collect nonsensitive failure patterns (e.g. hook errors) for improvement.
    • Ensure all telemetry is optin and redacts secrets.

Constraints

  • Must remain EUPL1.2.
  • Must preserve ethics modal and blueteam posture.
  • Avoid shell execution where possible in Gemini CLI.