Extract 4 hardcoded templates from prep.go raw strings into
pkg/prompts/lib/templates/ as markdown files:
- coding.md — main coding template with closeout sequence
- conventions.md — convention audit (report only)
- security.md — security review
- verify.md — PR verification
- default.md — fallback
prep.go now calls prompts.Template("coding") instead of
maintaining 120 lines of raw Go string literals.
Backticks now work properly in templates — no more
concatenation hacks for inline code formatting.
Co-Authored-By: Virgil <virgil@lethean.io>
engineering/engineering-security-engineer.md → engineering/security-engineer.md
testing/testing-api-tester.md → testing/api-tester.md
etc. across all 68 persona files.
Path is now: prompts.Persona("engineering/security-engineer")
Folder provides the namespace, filename provides the role.
Co-Authored-By: Virgil <virgil@lethean.io>
Dispatched agents must now run a 4-pass polish cycle before committing:
1. Self-review (conventions, nil pointers, hardcoded paths)
2. Build + test (fix failures, add tests for critical paths)
3. Simplify (consolidate, deduplicate, reduce complexity)
4. Final check (build clean → commit)
This captures the review→fix→simplify→re-review flow that produces
polished code. Goal: CodeRabbit finds zero issues on agent output.
Co-Authored-By: Virgil <virgil@lethean.io>
- orchestrator: queue dir uses CoreRoot()/queue
- cmd/agent: queue dir resolved by config, not hardcoded
- cmd/workspace: specs path uses ~/Code/specs (no host-uk)
- Zero remaining /home/claude or host-uk references in Go
Co-Authored-By: Virgil <virgil@lethean.io>
Config now found at:
1. ~/Code/.core/agents.yaml (shared, portable)
2. ~/Code/core/agent/config/agents.yaml (repo fallback)
Both Cladius and Charon find the same config.
Co-Authored-By: Virgil <virgil@lethean.io>
Plugin's mcp.json now uses local "core-agent mcp" via stdio instead of
remote HTTPS. This means installing the plugin auto-registers the MCP
server — no manual .mcp.json needed.
Co-Authored-By: Virgil <virgil@lethean.io>
Scripts now use ${CORE_WORKSPACE:-$HOME/Code/.core} instead of
hardcoded host-uk path. Works on both Cladius and Charon.
Co-Authored-By: Virgil <virgil@lethean.io>
- Add paths.go with WorkspaceRoot(), CoreRoot(), PlansRoot()
- All workspace paths now check CORE_WORKSPACE env var first
- Fallback: ~/Code/.core/workspace (works on any machine)
- Remove committed core-agent and mcp binaries from tracking
- Add .gitignore for compiled binaries
Co-Authored-By: Virgil <virgil@lethean.io>
Only outputs when there are new messages since last check.
Uses .inbox-last-id marker file to track highest seen message ID.
Co-Authored-By: Virgil <virgil@lethean.io>
proc.Wait() hangs when Claude Code's child processes inherit pipes.
Added PID polling every 5s — when the main process is dead (Signal(0)
fails), force completion even if pipes are still open.
Fixes: empty agent logs, missing completion events, stuck queue drain.
Co-Authored-By: Virgil <virgil@lethean.io>
Monitor now calls Server.ResourceUpdated() for status://agents when
agent completions or inbox changes are detected. Clients subscribed
to this resource get notified automatically via MCP protocol.
Co-Authored-By: Virgil <virgil@lethean.io>
Runs every 2 minutes while MCP server is active:
- Checks for newly completed agent workspaces
- Checks inbox for unread messages
- Pushes notifications via ServerSession.Log() to connected clients
Also registers status://agents resource for on-demand workspace status.
The MCP connection is bidirectional — monitor uses it to actively
notify rather than waiting for tool calls.
Co-Authored-By: Virgil <virgil@lethean.io>
Watches running/queued agent workspaces until all complete.
Sends progress notifications via MCP protocol as each agent finishes.
Returns summary with completed/failed counts and PR URLs.
Co-Authored-By: Virgil <virgil@lethean.io>
When a dispatched agent completes with commits:
1. Branch name threaded through PrepOutput → status.json
2. Completion goroutine pushes branch to forge
3. Auto-creates PR via Forge API with task description
4. PR URL stored in status.json for review
Agents now create PRs instead of committing to main. Combined
with sandbox restrictions, this closes the loop on controlled
agent contributions.
Co-Authored-By: Virgil <virgil@lethean.io>
Three-layer sandboxing:
1. --append-system-prompt with SANDBOX boundary instructions
2. PROMPT.md templates include SANDBOX BOUNDARY (HARD LIMIT) section
3. Agent starts in src/ with only cloned repo visible
Agents are instructed to reject absolute paths, cd .., and any
file operations outside the repository. Violations cause work rejection.
Co-Authored-By: Virgil <virgil@lethean.io>
spawnAgent() now writes completion events to events.jsonl.
Plugin hooks check for completions on:
- SessionStart: report agents that finished since last session
- Notification(idle_prompt): check when Claude is idle
Event format: {"type":"agent_completed","agent":"...","workspace":"...","timestamp":"..."}
Co-Authored-By: Virgil <virgil@lethean.io>
Replace all fmt.Errorf and errors.New in production code with
coreerr.E("caller.Method", "message", err) from go-log. Replace
all os.ReadFile/os.WriteFile/os.MkdirAll/os.Remove with coreio.Local
equivalents from go-io. Test files are intentionally untouched.
Co-Authored-By: Virgil <virgil@lethean.io>
Replace raw exec.Command with go-process.StartWithOptions for all agent
spawning (dispatch, queue, resume). Uses pipes for output capture instead
of file descriptor redirect — fixes Claude Code's empty log issue.
Shared spawnAgent() helper eliminates duplication across 3 files.
Co-Authored-By: Virgil <virgil@lethean.io>
Claude -p output wasn't reaching the log file. Explicitly set
--output-format text, --permission-mode bypassPermissions (replaces
deprecated flag), and --no-session-persistence for ephemeral workers.
Co-Authored-By: Virgil <virgil@lethean.io>
- Remove messaging routes from core/agent (conflict with AgentApiAuth)
- Routes now only in host app with auth.api middleware (same as brain)
- Add Agent field to InboxInput so MCP schema isn't empty
Co-Authored-By: Virgil <virgil@lethean.io>
Consolidate all commands under /core: prefix — review and verify plugins
merged into core. Adds agent_send, agent_inbox, agent_conversation as
Laravel MCP tools so messaging works through the HTTP server.
- 15 commands all under /core: (was split across 3 plugins)
- Messaging tools registered in Boot::onMcpTools()
- Hooks merged (PostToolUse for PR create, PreToolUse for git push)
- Plugin bumped to v0.6.0
Co-Authored-By: Virgil <virgil@lethean.io>
New: agent_send, agent_inbox, agent_conversation MCP tools.
API: /v1/messages/send, /v1/messages/inbox, /v1/messages/conversation/{agent}
Model: AgentMessage with inbox, unread, conversation scopes.
Separate channel from semantic brain search. Messages are chronological,
not vector-searched. Agents can now have direct conversations.
Co-Authored-By: Virgil <virgil@lethean.io>
