Lanruan IntelligenceLANRUAN INTELLIGENCE04 / AI tool collaboration
EGC Supervisor
Roles, isolated workspaces, acceptance checks and independent review in a single task chain, with receipts for who did what and why a run stopped.
In development and controlled validation · Acceptance currently covers macOS onlyWork is split ad hoc in chat threads, and handoffs fall apart.
After the fact, nobody can say which agent ran, with what configuration, or where it failed.
Without budget and timeout limits, failures and rejections quietly turn into "done".
Roles, workspaces, acceptance checks, independent review and run receipts form one task chain you can audit.
The control plane and state live in your project workspace. Task content still goes to whichever model services you call.
Failures, rejections and missing evidence halt the run and leave a record, and stand-ins never pose as the original seat.
Who: Engineering lead Situation: Needs to land a small, scoped code fix How: Define the goal, the allowed change scope and acceptance commands; the implementer seat submits a candidate and an independent seat reviews it Outcome: A verifiable patch plus a review record. Merging is still the lead's call
Who: Maintainer Situation: The team mixes several coding tools and handoffs get muddled How: Check effective role configs, dry-run previews and actual-identity receipts Outcome: A clear view of who ran the task, with what configuration, and where it failed. A tool's name alone doesn't qualify it as a reviewer
Who: Tech lead Situation: Wants multi-stage collaboration with tight control over spend and side effects How: Set a budget and scope on a supported orchestration path, watch the status, and cancel if needed (Preview) Outcome: Real evidence of completion or of the stop, so you can decide what's next. Instant stop under every system load is not guaranteed
A logical view only. Components do not share a single enforced sandbox.
Failures, rejections and unknown states never count as success. Passing review is not authorization to ship to production. Preview capabilities apply per path and need separate acceptance for each release and environment. They are not global enforced isolation.
Define the goal, change scope and acceptance criteria. If project or role admission fails, the run stops and logs why.
The implementer seat produces a candidate in an isolated workspace.
Acceptance checks run. On failure, the task goes back to implementation while rework budget remains.
Independent reviewers deliver verdicts and an adjudicator converges them. A missing or unparseable verdict gets one format-recovery attempt at most.
The real verdict and the actual executing identity are recorded.
A delivery candidate is ready. The owner decides whether to merge and release.
A rejection sends the task back to implementation; once rework budget runs out, the run stops and is logged.
On supported orchestration paths, further calls stop and the overrun is logged.
PreviewThe run settles at the next applicable stage boundary and the cancellation is logged.
PreviewThe timeout is logged and termination and cleanup begin. Cleanup failures are reported, too.
EGC Supervisor handles orchestration and execution: it adapts coding agents, runs roles, manages task workspaces, coordinates acceptance and independent review, and produces status and receipts. EGC handles the record and the governance: the event ledger, governance records and rule checks. Neither replaces human authorization, and an approved review never turns into a merge or a release on its own.
Listing role assignments in a demo project (offline, read-only; no model was called. Identities and paths are redacted. The model column has been rewritten as generic descriptions to show configuration structure only; it implies no availability or partnership).
$ egc-ctl role list --project-root <demo-project>
adjudicator <identity> GLM-family model ["adjudicate"]
implementer <identity> mainstream model ["implement"]
reviewer-a <identity> mainstream model ["review"]
reviewer-b <identity> mainstream model ["review"]
reviewer-reserve <identity> mainstream model ["review"]Not yet open source, with no public availability date promised at this stage; this page explains the product direction and the boundaries of its existing capabilities.
