Runner now creates a reservation entry (PID=-1) in the workspace Registry
immediately when approving a dispatch. This prevents parallel requests
from all seeing count < limit before any spawn completes.
Reservations are counted by countRunningByAgent/ByModel (PID < 0 = always
count). Agentic overwrites with real PID via TrackWorkspace after spawn.
Co-Authored-By: Virgil <virgil@lethean.io>
Moves concurrency, queue drain, workspace lifecycle, and frozen state
from agentic/prep into pkg/runner/ — a standalone Core service that
communicates via IPC Actions only.
- runner.Register wires Actions: dispatch, status, start, stop, kill, poke
- runner.HandleIPCEvents catches AgentCompleted → ChannelPush + queue poke
- Agentic dispatch asks runner for permission via c.Action("runner.dispatch")
- Dispatch mutex moved to struct-level sync.Mutex (fixes core.Lock init race)
- Registry-based concurrency counting replaces disk scanning
- TrackWorkspace called on both queued and running status writes
- SpawnQueued message added for runner→agentic spawn requests
- ChannelPush message in core/mcp enables any service to push channel events
- 51 new tests covering runner service, queue, and config parsing
Co-Authored-By: Virgil <virgil@lethean.io>
- cloneWorkspaceDeps: reads go.mod, clones Core ecosystem modules from Forge
into workspace alongside ./repo, rebuilds go.work with all use directives
- Deduplicates deps (dappco.re + forge.lthn.ai map to same repos)
- Container chmod: workspace files made writable before exit so host can clean up
- GONOSUMCHECK for local workspace modules (bypass checksum for dev branches)
- Removed stale OLLAMA_HOST env from container
Co-Authored-By: Virgil <virgil@lethean.io>
The monitoring goroutine that waits for agent completion was a raw `go func()`
that Core didn't know about. ServiceShutdown killed it immediately on CLI exit.
Now uses PerformAsync which registers with Core's WaitGroup:
- ServiceShutdown waits for all async tasks to drain
- `core-agent workspace dispatch` blocks until agent completes
- Agent lifecycle properly tracked by the framework
Also whitelist agentic.monitor.* and agentic.complete in entitlement checker.
Co-Authored-By: Virgil <virgil@lethean.io>
The global process.StartWithOptions() requires process.SetDefault() which
was never called. Use core.ServiceFor[*process.Service] to get the registered
service instance directly — same code path, proper Core wiring.
Fixes: dispatch failing with "failed to spawn codex" on CLI dispatch.
Co-Authored-By: Virgil <virgil@lethean.io>
- proc.go: ensureProcess() as temporary bridge until go-process gets v0.7.0 update
- processIsRunning/processKill: use go-process ProcessID when available, fall back to PID
- WorkspaceStatus: add ProcessID field for go-process managed lookup
- dispatch.go: simplified spawnAgent goroutine — uses proc.Done() instead of syscall poll
- Removed syscall import from dispatch.go
Next: update go-process to v0.7.0 Core contract, then replace
syscall.Kill calls in queue.go, shutdown.go, status.go, dispatch_sync.go
with processIsRunning/processKill.
Coverage: 78.1%, 802 tests
Co-Authored-By: Virgil <virgil@lethean.io>
Services are now registered during Core construction:
core.New(
core.WithService(agentic.Register),
core.WithService(monitor.Register),
core.WithService(brain.Register),
)
- Remove initServices() closure — services created once in conclave
- Commands use c.ServiceStartup()/c.ServiceShutdown() for lifecycle
- Service instances retrieved via c.Config() for MCP tool registration
- run/orchestrator reduced to ServiceStartup + block + ServiceShutdown
- run/task uses conclave's agentic instance
Co-Authored-By: Virgil <virgil@lethean.io>
Phase 3 of Core DI migration:
- Remove CompletionNotifier interface from pkg/agentic
- dispatch.go emits messages.AgentStarted/AgentCompleted via c.ACTION()
- monitor registers IPC handlers in SetCore() — handleAgentStarted/handleAgentCompleted
- Remove circular callback wiring (SetCompletionNotifier) from main.go
- Export ReadStatus for cross-package use
- Update run/orchestrator to use SetCore() instead of SetCompletionNotifier()
Services now communicate through typed messages, not direct references.
Co-Authored-By: Virgil <virgil@lethean.io>
- Export ReadStatus (was readStatus) for cross-package use
- AgentCompleted now emits agent.completed with repo/agent/workspace/status
for every finished task, not just failures
- queue.drained only fires when genuinely empty — verified by checking
PIDs are alive via kill(0), not just trusting stale status files
- Fix Docker mount paths: /root/ → /home/dev/ for non-root container
- Update all callers and tests
Co-Authored-By: Virgil <virgil@lethean.io>
- Add background queue runner (runner.go) — 30s tick + poke on completion
- drainQueue now loops to fill all slots per tick
- Add run orchestrator command — standalone queue runner without MCP
- Slim agentic_status — stats only, blocked workspaces listed
- Docker containerised dispatch — all agents run in core-dev container
- Forge stopwatch start/stop on issue when agent starts/completes
- issue create supports --milestone, --assignee, --ref
- Auto-PR targets dev branch (not main)
- PR body includes Closes #N for issue-linked work
- CLI usage strings use spaces not slashes
- Review agent uses exec with sandbox bypass (not codex review subcommand)
- Local model support via codex --oss with socat Ollama proxy
Co-Authored-By: Virgil <virgil@lethean.io>
CompletionNotifier interface now has AgentStarted() and
AgentCompleted() instead of Poke(). Dispatch pushes notifications
directly to monitor with agent/repo/status data. Monitor pushes
MCP channel events immediately — no scanning, no dedup maps,
no filesystem polling latency.
Events.jsonl kept as audit log only, not notification mechanism.
Timer-based scan kept for startup seeding and stale detection.
Co-Authored-By: Virgil <virgil@lethean.io>
- emitStartEvent fires when agent spawns (dispatch.go)
- Monitor detects new "running" workspaces and pushes agent.started
channel notification with repo and agent info
- agent.complete already included blocked/failed status — no change
- Both old and new workspace layouts supported
Co-Authored-By: Virgil <virgil@lethean.io>
When model variant was specified, the splice inserted --model
between -o and its value, making Codex see -o with no file path.
Fixed by appending --model and prompt sequentially instead of
splicing into the middle of the args slice.
Co-Authored-By: Virgil <virgil@lethean.io>
Agent name like "codex:gpt-5.3-codex-spark" contains a colon which
breaks file paths. Use base name (before ":") for the log file.
Co-Authored-By: Virgil <virgil@lethean.io>
After agent completes, run build + vet + test before creating PR.
If QA fails, mark workspace as failed with "QA check failed" —
bad code never gets PR'd.
Supports Go (build/vet/test), PHP (composer install/test),
and Node (npm install/test). Unknown languages pass through.
Co-Authored-By: Virgil <virgil@lethean.io>
Major simplification of the dispatch model:
- Workspace dir: .core/workspace/{org}/{repo}/{pr|task|branch|tag}/
- Clone into repo/ (not src/), metadata in .meta/
- One of issue, pr, branch, or tag required for dispatch
- All context (brain, consumers, git log, wiki, plan) assembled
into prompt string — no TODO.md, PROMPT.md, CONTEXT.md files
- Resume detection: skip clone if repo/.git exists
- Default agent changed to codex
- spawnAgent drops srcDir param, runs from repo/
- No --skip-git-repo-check (repo/ IS a git repo)
- All downstream files: srcDir → repoDir
Track PRs, not workspace iterations.
Co-Authored-By: Virgil <virgil@lethean.io>
