AnswersPublished16 sources
Keep Pi for Now. Fix the Execution Chain First.
The available evidence points to an incomplete workflow, not a demonstrated failure of Pi or proof that the system has suffered an actual split in authority. The proposed direction keeps one Python controller in charge, with separate planning, research, coding and review roles—and requires a person to explicitly aut...
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- The available evidence points to an incomplete workflow, not a demonstrated failure of Pi or proof that the system has suffered an actual split in authority.
- The proposed direction keeps one Python controller in charge, with separate planning, research, coding and review roles—and requires a person to explicitly authorize each run.
- Cline’s documented CLI and planning features are not enough to establish that a full migration would solve the current blockers.
- The shortest path forward is to connect and test a small end to end run before adding more tooling, Jev extensions or interface polish.
![[GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you operate via client side tool execution on the](https://d15bya8fi85z1v.cloudfront.net/chat/VjcFMwsvhIToc7jsPOzIkYyn4yG3/images/ED6940EF4A064E923A15/EF6722A07838CE4C275B.png)
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: [GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you operate via client side tool execution on the. Article summary: <tool call {"name":"attempt completion","arguments":{"result":"建议保留现有 V5 的 Python 控制器和 Pi 执行适配，改为“人显式启动一次、控制器唯一裁决、四槽位分权、IDE 只读观察”；暂不把迁移 Cline 作为解阻前提。当前证据支持“链路尚未接完整且门禁未实例化”，不支持“Pi 底层选型已经失败”，也不能证明已发生实际脑裂。\n\n 1.. Topic tags: general web, ai safety, llm, ai, automation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake n
A coding-agent workflow can have plenty of components and still lack a reliable path from “start” to “done.” That is the central finding of a review of a proposed multi-agent setup: the immediate problem is not evidence that Pi is the wrong coding agent. It is that key parts of the workflow have not yet been connected and verified.
The recommendation is to keep the existing Python controller and Pi execution adapter for now. Make one controller authoritative, require a person to explicitly authorize a run, separate coding from approval, and treat the IDE as a way to observe work—not as the process that keeps it running.
What the current evidence shows—and what it doesn’t
The implementation plan is still marked BLOCKED, while its associated architecture decision record (ADR) remains PROPOSED. A structured checklist records seven historical items as complete and one as suspended, but that is not enough to calculate the completion rate of the project as a whole.
A real-run task is still suspended with no receipt. The latest historical run record describes a network error and an unfinished task. That is evidence of a blocked network path; it does not identify whether the cause was Pi, a gateway or an upstream service.
Logs in the engineering directory record successful checks for checkpoint wiring, the supervisor and incident handling. But this review only read those existing logs; it did not rerun the checks against the current code. Meanwhile, the main loop blocks completion and the next task because the Git transaction has not been connected. The Git transaction module already exists, so the issue is not that there is no module—it is that it is not wired into the main execution path.
Other gaps are more specific: the model registry has only two candidates and no default for coding; the router has no separate review action; and Pi is launched with extensions disabled. That means the current setup cannot also claim that pi-jev is loaded.
Taken together, the evidence supports a narrow conclusion: the chain is incomplete and several proposed safeguards have not yet been turned into executable acceptance checks. It does not establish that Pi has failed as a platform or that the system has already experienced a real authority split.
Why a Cline migration is not the first fix
Cline’s documentation describes CLI automation, auto-approval and a plan mode for read-only exploration. 1
2
3 Those capabilities are relevant, but the reviewed evidence does not establish the broader set of background Hub, shared-session and lifecycle guarantees described in the proposal. So Cline has not been shown to pass every required gate—and a full migration would add new protocols, configuration, permissions and regression work before it is clear that it addresses the present blockers.
The conditional first choice is therefore the existing Python controller with Pi in four distinct roles. Cline could remain a non-critical-path option for planning or research if it proves useful in practice. It should not become a second scheduler with its own authority over task progress.
The design goal is four role-specific permissions, not four always-on agents. Different models can provide independent judgments, while deterministic software controls state changes. Those are different jobs.
One controller, four roles, one human authorization
The proposed roles are:
- Planning: Opus or Astra explores and narrows the plan after a person provides a starting point. It may update the shared plan or proposal only when authorized; it cannot start coding.
- Research: Astra Academic gathers sources and evidence when needed. It does not edit business code or change acceptance criteria.
- Coding: Pi with Sonnet works in a fresh session for each task or outer-loop repair. It can investigate, edit and diagnose, but cannot approve its own work or mark a task complete.
- Review: An independent GPT-6-Sol session checks a frozen candidate against the task’s acceptance criteria, relevant source code and outer-loop verification receipt. It does not read the coding conversation or modify business code.
These are proposed role labels, not verified model identifiers. The existing registry lists GPT-6.1-Sol, so its relationship to the requested GPT-6-Sol name still needs to be reconciled.
The Python controller remains the sole scheduler and authority over task state. Four tmux slots, if used, are places to run or observe roles—not separate queues, and not a requirement to keep four models online. The task list bound to the plan is the source of progress. A status card should reflect that list, not become another independent progress ledger.
Crucially, a ready plan, a successful review, a change to the entry point or a reloaded IDE must not start coding by themselves. A person must explicitly authorize a defined task set, scope and budget. That authorization applies to a particular run; submitting it again should return the existing run reference, not create a second writer.
Once authorized, the workflow can automatically research, code, verify, review, checkpoint and move to the next task within the approved scope. A change in scope or acceptance criteria, an expired authorization or a failed key assumption should send control back to the person.
The IDE is for viewing changes and results over SSH. If someone needs to edit the same workspace during a run, the run should first be paused and cleaned up; after the edit, the controller can be handed control again.
Jev can assist with judgment, but it should not grant authority
There are several different Jev extensions, and their capabilities should not be treated as interchangeable. The reviewed references describe tools for typed decisions, tool-call judgment and output review, with different behavior when credentials or services are unavailable. 9
10
11
12 That is not enough to assume that one extension’s enforcement switch also covers timeouts, exhausted credits or other service failures.
Keep the four role assignments fixed rather than letting Jev freely switch a session between planning, coding and review. Separate model-routing advice from a security gate, too: a routing failure can use a pre-authorized fixed model, but a safety gate declared mandatory should block on failure rather than silently allow the action.
Jev may help assess scope, risk or a failure category. Programmatic controls must still enforce path permissions, account separation, prohibited operations, timeouts and budgets. Prompts and semantic judgments are not substitutes for those boundaries.
Jev need not block the first fixed-model end-to-end test. But if the intended deliverable includes Jev, the combined setup needs its own acceptance test; a fixed-model test is not a substitute.
The shortest route to a verifiable workflow
- Clarify the decision in the existing ADR. Add an explicit replacement for any default design in which a front-end agent launches a one-off model run. State the boundaries: human authorization to start, one controller for task state, separate coding and review roles, and read-only observation as the default IDE role.
- Turn requirements into runnable acceptance checks. The current structured checklist mainly describes Python unit tests. A real disconnection, model review or Git transaction should not be represented as passing merely because a unit test passed; specify how each check is tied to a real run receipt.
- Test the smallest real chain. Use one fixed coding model and an independent reviewer on two small, dependent tasks. Verify a failure, a repair in a fresh session, outer-loop checks, blind review, a checkpoint and progression to the next task. Capture events after the client disconnects. Do this before building a four-pane interface or dynamic routing.
- Fix the Git transaction for dependent tasks. The existing module supports ordinary repositories but prohibits overlap between task-managed paths. Since a later task may need to change a file touched by an earlier one, the transaction needs a per-task, per-checkpoint handoff—not a rule that every task must own disjoint files.
- Add research, Jev and role displays only after the core chain works. Rerun checks that are relevant to changed interfaces and evidence. A saved state, unrelated document edit or status query should not trigger unnecessary business-logic retesting. Seek ADR acceptance only after the full chain and interruption points pass.
A worktree migration is not recommended as part of this immediate unblock. It would change both the shared-workspace arrangement and the ordinary-repository transaction interface at once. First establish a single-writer chain in an isolated coding workspace; consider a different repository topology as a separate decision.
This is an architecture review and proposed sequence—not a claim that the changes have been implemented. No plan, ADR or production files were modified, and no tests, terminal commands, live model runs or background tasks were started. The practical takeaway is straightforward: don’t replace Pi just to compensate for missing connections. First make authorization, task state, independent review and Git checkpoints part of one small, verifiable execution chain.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Keep Pi for Now. Fix the Execution Chain First."?
The available evidence points to an incomplete workflow, not a demonstrated failure of Pi or proof that the system has suffered an actual split in authority.
What are the key points to validate first?
The available evidence points to an incomplete workflow, not a demonstrated failure of Pi or proof that the system has suffered an actual split in authority. The proposed direction keeps one Python controller in charge, with separate planning, research, coding and review roles—and requires a person to explicitly authorize each run.
What should I do next in practice?
Cline’s documented CLI and planning features are not enough to establish that a full migration would solve the current blockers.