Inline instructions for Codex agents — no spec references,
just the actual rules with code examples. Extracted into every
workspace automatically.
Co-Authored-By: Virgil <virgil@lethean.io>
Reference files (.core/reference/) are now part of the embedded
workspace template. ExtractWorkspace extracts them automatically —
no hardcoded filesystem paths, ships with the binary.
Co-Authored-By: Virgil <virgil@lethean.io>
Parallel to the Claude Code plugin (claude/), core-agent now supports
Codex with full feature parity:
- .codex/config.toml — model, profiles (review/quick/implement/lem),
MCP server, local model providers (Ollama/LM Studio)
- .codex/agents/ — reviewer, migrator, fixer (AX-aware)
- .codex/rules/ — Starlark rules for sandbox control
- .agents/skills/ — 9 skills matching Claude plugin
- AGENTS.md — Codex project instructions (like CLAUDE.md)
Supports --oss mode for local models (LEM via Ollama).
Same binary, two entry points.
Co-Authored-By: Virgil <virgil@lethean.io>
Replace separate go-io (coreio) and go-log (coreerr) packages with
Core's built-in Fs and error/logging functions. This is the reference
implementation for how all Core ecosystem packages should migrate.
Changes:
- coreio.Local.Read/Write/EnsureDir/Delete/IsFile → core.Fs methods
- coreerr.E() → core.E(), coreerr.Info/Warn/Error → core.Info/Warn/Error
- (value, error) return pattern → core.Result pattern (r.OK, r.Value)
- go-io and go-log moved from direct to indirect deps in go.mod
- Added AX usage-example comments on key public types
- Added newFs("/") helper for unrestricted filesystem access
Co-Authored-By: Virgil <virgil@lethean.io>
inbox.message events now include full message objects (id, from,
subject, content) so recipients can read messages inline without
calling agent_inbox. Charon's messages arrive directly in session.
Co-Authored-By: Virgil <virgil@lethean.io>
Both checkCompletions and checkInbox now seed on first run —
existing workspaces and messages are recorded without firing
channel events. Only genuinely new events trigger notifications.
Tests updated to pre-seed flags and filter debug events.
Co-Authored-By: Virgil <virgil@lethean.io>
Track inbox by highest message ID instead of unread count. Fixes:
- API pagination limit (max 20) no longer causes missed notifications
- Restart no longer floods with all existing unread messages (seeded)
- Each new message fires exactly once regardless of read state
Added MONITOR_INTERVAL env override and debugChannel helper for
faster iteration during channel development.
All three channel types confirmed working:
- agent.complete: workspace status changes
- inbox.message: new messages by ID tracking
- monitor.debug: real-time debug trace
Co-Authored-By: Virgil <virgil@lethean.io>
The MCP agent_inbox tool wraps the response as {messages:[...]},
but the raw API returns {data:[...]}. The monitor calls the raw
API directly, so it needs to parse {data:[...]}.
Verified with curl against live API. Removed debug channel events.
Co-Authored-By: Virgil <virgil@lethean.io>
New plugin: claude/research (core-research v0.1.0)
- 12 collection/research skills moved from core
- collection subsystem (hooks, scripts) moved
- Installed separately when research work is needed
Core plugin bumped to v0.12.0 — now focused on:
- Dispatch, review, messaging, OpenBrain (operational)
- Deploy, app-split, repo-sweep (infrastructure)
- Architecture review, security review (quality)
Marketplace updated with CoreResearch entry.
Co-Authored-By: Virgil <virgil@lethean.io>
Critical: /core:remember rewritten to use brain_remember MCP tool
instead of missing capture-context.sh script.
Major: /core:pipeline persona paths fixed — now reference
pkg/lib/persona/ with correct slugs instead of nonexistent
agents/engineering/ directory.
Major: Added YAML frontmatter (name + description) to all 12
collection skill SKILL.md files for Claude Code auto-discovery.
Removed empty channel/ directory.
Co-Authored-By: Virgil <virgil@lethean.io>
High: verify rebase push now targets Forge remote, not local origin
High/Security: planPath sanitises ID via filepath.Base to prevent
path traversal in plan read/update/delete
Medium: mirror compares and pushes local default branch, not HEAD
Findings 3-6 verified as false positives/known issues (bridge async
by design, API returns top-level memories not {data:...}, inbox uses
{messages:...} confirmed against live API).
Co-Authored-By: Virgil <virgil@lethean.io>
Medium: remote dispatch now propagates inner dispatchOut.Success
instead of hardcoding true.
Low: updateStatus clears stale question field when new state
has no question, preventing leftover rejection messages.
Co-Authored-By: Virgil <virgil@lethean.io>
High/Security: sanitise input.Repo via filepath.Base to prevent
path traversal in workspace prep (../escape from CODE_PATH).
High/Security: sanitise repo.Repo from API response in syncRepos
to prevent path traversal via crafted checkin responses.
Medium: dispatchFixFromQueue now returns error, review_queue checks
success before recording fix_dispatched.
Known issues updated with async bridge provider findings.
Co-Authored-By: Virgil <virgil@lethean.io>
API enhancements, test coverage gaps, conventions, and compile
issues — all acknowledged and tracked for future work.
Co-Authored-By: Virgil <virgil@lethean.io>
High: workspace names use UnixNano to prevent same-second collisions
High: sync only pulls the branch the server reported (was pulling current)
Medium: drainQueue serialised via mutex to prevent concurrent over-dispatch
Medium: remote_status checks JSON-RPC error field before reporting success
Medium: dead agent PIDs without output log marked failed, not completed
Low: detectLanguage uses ordered slice instead of map for deterministic results
Also: URL-encoded agent names in messaging, monitor inbox, and sync endpoints.
Co-Authored-By: Virgil <virgil@lethean.io>
- harvest: message says 'ready-for-review' not 'pushed'
- sync: timestamp advanced after pulls, not before
- sync: accepts main/master/reported branch, not just main
- inbox: checks CORE_BRAIN_KEY env before falling back to file
- inbox: parses 'from' not 'from_agent', 'messages' not 'data'
- queue: strips variant suffix for rate limit lookup (claude:opus → claude)
- review_queue: respects ReviewQueueInput.Reviewer instead of hardcoding coderabbit
- tests: updated to match real API response structure
Co-Authored-By: Virgil <virgil@lethean.io>
High: harvest no longer hardcodes 'main' — detects default branch
via symbolic-ref/rev-parse fallback. Repos with master/other
default branches are now harvested correctly.
High: empty task no longer produces invalid 'agent/' branch name.
Falls back to issue-N or work-timestamp. Branch creation errors
are now surfaced instead of silently ignored.
High: PHP verification no longer returns passed:true when no test
runner exists. Untested PHP repos correctly fail verification.
(brain/direct.go findings 5-6 verified as false positives — API
returns top-level keys, not {data: ...} envelope)
Co-Authored-By: Virgil <virgil@lethean.io>
High: Fix missed-notification bug — track completions by workspace
name instead of count, so harvest status rewrites don't suppress
future notifications. Also tracks blocked/failed terminal states.
High: Safety gate fail-closed — check ALL changed files (not just
added), reject on git diff failure instead of proceeding.
Medium: emitCompletionEvent now passes actual status (completed,
failed, blocked) instead of hardcoding "completed".
Medium/AX: Harvest no longer auto-pushes to source repos. Sets
status to ready-for-review only — pushing happens during explicit
review, not silently in the background.
Co-Authored-By: Virgil <virgil@lethean.io>
