core/agent

Snider 90b03191b2 feat(agent): v0.2.0 — HTTP daemon, remote dispatch, review queue, verify+merge

Major additions:
- core-agent serve: persistent HTTP daemon with PID file, health check, registry
- agentic_dispatch_remote: dispatch tasks to remote agents (Charon) over MCP HTTP
- agentic_status_remote: check remote agent workspace status
- agentic_mirror: sync Forge repos to GitHub mirrors with file count limits
- agentic_review_queue: CodeRabbit/Codex review queue with rate-limit awareness
- verify.go: auto-verify (run tests) + auto-merge + retry with rebase + needs-review label
- monitor sync: checkin API integration for cross-agent repo sync
- PostToolUse inbox notification hook (check-notify.sh)

Dispatch improvements:
- --dangerously-skip-permissions (CLI flag changed)
- proc.CloseStdin() after spawn (Claude CLI stdin pipe fix)
- GOWORK=off in agent env and verify
- Exit code / BLOCKED.md / failure detection
- Monitor poke for instant notifications

New agent types:
- coderabbit: CodeRabbit CLI review (--plain --base)
- codex:review: OpenAI Codex review mode

Integrations:
- CODEX.md: OpenAI Codex conventions file
- Gemini extension: points at core-agent MCP (not Node server)
- Codex config: core-agent MCP server added
- GitHub webhook handler + CodeRabbit KPI tables (PHP)
- Forgejo provider for uptelligence webhooks
- Agent checkin endpoint for repo sync

Co-Authored-By: Virgil <virgil@lethean.io>

2026-03-17 17:45:04 +00:00

2.4 KiB

Raw Blame History

CODEX.md

Instructions for OpenAI Codex when working in the Core ecosystem.

MCP Tools Available

You have access to core-agent MCP tools. Use them:

brain_recall — Search OpenBrain for context about any package, pattern, or decision
brain_remember — Store what you learn for other agents (Claude, Gemini, future LEM)
agentic_dispatch — Dispatch tasks to other agents
agentic_status — Check agent workspace status

ALWAYS brain_remember significant findings — your deep analysis of package internals, error patterns, security observations. This builds the shared knowledge base.

Core Ecosystem Conventions

Go Packages (forge.lthn.ai/core/*)

Error handling: coreerr.E("pkg.Method", "what failed", err) from go-log. NEVER fmt.Errorf or errors.New.
- Import as: coreerr "forge.lthn.ai/core/go-log"
- Always 3 args: operation, message, cause (use nil if no cause)
- coreerr.E returns *log.Err which implements error and Unwrap()
File I/O: coreio.Local.Read/Write/Delete/EnsureDir from go-io. NEVER os.ReadFile/WriteFile/MkdirAll.
- Import as: coreio "forge.lthn.ai/core/go-io"
- Security: go-io validates paths, prevents traversal
Process management: go-process for spawning external commands. Supports Timeout, GracePeriod, KillGroup.
UK English: colour, organisation, centre, initialise (never American spellings)
Test naming: TestFoo_Good (happy path), TestFoo_Bad (expected errors), TestFoo_Ugly (panics/edge cases)
Commits: type(scope): description with Co-Authored-By: Virgil <virgil@lethean.io>

PHP Packages (CorePHP)

Actions pattern: Single-purpose classes with use Action trait, static ::run() helper
Tenant isolation: BelongsToWorkspace trait on ALL models with tenant data
Strict types: declare(strict_types=1) in every file
Testing: Pest syntax, not PHPUnit

Review Focus Areas

When reviewing code, prioritise:

Security: Path traversal, injection, hardcoded secrets, unsafe input
Error handling: coreerr.E() convention compliance
File I/O: go-io usage, no raw os.* calls
Tenant isolation: BelongsToWorkspace on all tenant models (PHP)
Test coverage: Are critical paths tested?

Training Data

Your reviews generate training data for LEM (our fine-tuned model). Be thorough and structured in your findings — every observation helps improve the next generation of reviews.

2.4 KiB Raw Blame History