AnswersPublished12 sources
A leaner AI-agent stack starts by knowing what to stop building
The review recommends ending development on V5’s execution layer, while keeping its analysis methods, decision records and historical evidence. Its proposed V6 stack has one orchestrator, one execution framework, external validation and a small, evidence based experience file.
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- The review recommends ending development on V5’s execution layer, while keeping its analysis methods, decision records and historical evidence.
- Its proposed V6 stack has one orchestrator, one execution framework, external validation and a small, evidence based experience file.
- Cline’s compatibility with the orchestrator, independent review isolation and reliable operation after a client disconnect remain unproven.
![[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/U0oDXR2yOIV8X9qOFcgLiFJcoUp2/images/A90C0EC23DEF257BE45D/1FF656F82B0EAD4E36E8.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":"建议冻结 V5 执行层，完整保留分析层；V6 条件采用“上游编排器＋单一 Cline 执行框架＋外部验证＋文本经验”。但当前不能宣布 V6 已可启动：Cline 生命周期适配、独立盲审隔离、Jev 外部门禁和真实断开闭环仍缺运行证据。\n\n BMRalph V6 目标架构方案与 V5 历史病理解剖报告\n\n . Topic tags: general web, llm, ai, workflow, code. 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 fa
A proposed redesign for BMRalph V6 starts with a restraint: stop adding machinery to V5. The review recommends freezing the older execution layer, retaining its analysis practices and historical records, and treating V6 as a new effort—not a continuation that inherits a green light from V5’s tests.
The proposed direction is deliberately small: let bmad-loop coordinate work, use Cline as the sole agent framework, keep validation outside the coding agent, and record only useful, evidence-backed lessons for later tasks.
That is an architectural recommendation, not a claim that the new stack is ready to run. The review flags unresolved questions about Cline’s integration, genuinely independent code review and whether a task can continue after the developer’s local client disconnects.
What the V5 review found—and what it did not
The review of the project files describes bmad_control_store.py as a 315-line, append-only file store—not a SQL database. It records decisions and checks such as file permissions, request versions and content fingerprints. But its documented boundary stops short of authenticating a person, granting execution authority, starting a task or recording task progress.
That distinction matters: a careful record of a decision is not the same thing as a completed work cycle. The review argues that V5 invested in a durable record-keeping mechanism before demonstrating a working path from authorization through execution, testing and completion.
It also corrects a claim about the module’s size: the current file inspected was 315 lines, not 1,728. The review says it found no evidence for the larger historical version or the commit associated with it. And while it warns that layers of locks and controls can impede progress, it does not establish that the store caused a deadlock. Its described lock is non-blocking: a conflict causes an operation to be refused rather than left waiting indefinitely.
The review likewise treats “context bloat” as unproven. Large code or records create a model-context burden only if they are repeatedly included in model input; the reviewed material did not establish that causal link.
Why passing unit tests is not a remote-workflow guarantee
The project plan records a historical result of 231 specialized tests passing. The review did not rerun them. More importantly, the same plan reportedly leaves real-backend behavior, independent review, work after a client closes and the full remote workflow unverified.
That does not make the tests worthless. Tests of simulated backends and local safeguards can show that particular components behave as intended. They cannot, by themselves, prove that a real agent can load a skill correctly, report completion reliably, survive a client disconnect or complete a review-and-commit cycle on a VPS.
The recommended accounting is therefore straightforward: keep the old tests as historical evidence, but do not treat them as V6 acceptance tests or as proof that V6 is ready.
One tool, one job
The proposed design draws firm boundaries between five components:
| Component | Intended responsibility | Boundary to protect |
|---|---|---|
| bmad-loop | Select an approved story, manage execution and verification, and handle retries and commits | Do not add a second orchestration loop around it |
| Cline CLI | Perform a single development or review task in a specified workspace | Do not turn it into a separate autonomous project manager |
| BMAD-METHOD | Supply roles and workflows when a task needs them | Do not copy the skills into a parallel private framework or load everything at once |
| Jev gate | Assess a bounded candidate change and help classify failures | Do not let it edit code, select work or overrule deterministic checks |
| Experience file | Preserve scoped lessons that may help future tasks | Do not let lessons change authorization or acceptance criteria |
Jev is described by its provider as a structured decision model: it evaluates a state and typed questions, then returns structured answers rather than acting as a coding or chat agent.1
2 That makes it a possible input to an external gate, not a substitute for ordinary checks such as verifying changed paths or running tests. The review recommends that an unknown or unavailable result should pause the workflow rather than count as approval.
The experience mechanism is similarly limited. The proposal borrows the broad idea of improving an execution harness through task-level refinement and lessons that carry forward, a theme explored in Harness² research.10 But it does not propose installing a full Harness² runner. Instead, it suggests a small text file containing a lesson, when it applies, and the evidence behind it. A failed task should not automatically become a supposedly proven rule.
The biggest unknown: can Cline satisfy the orchestrator’s contract?
The review’s compatibility concern is specific. It says bmad-loop’s documented built-in CLI list does not include Cline, while its general adapter route expects lifecycle hooks that the listed hook formats do not cover Cline. The installed Cline documentation was said to support features such as working directories and isolated state, but that is not the same as proving the complete integration.
Before V6 starts real work, a test must establish that a Cline session can be launched with a task, identified as a fresh session, reported as complete only when the task is genuinely complete, and stopped along with its tool processes when cancelled. If usage information is unavailable, the system must report it as unknown—not imply that a usage-based budget is enforced.
The design calls for separate development and review profiles, potentially using different models. But the proposed names “Sonnet 5.5” and “GPT-6-Sol” are targets in the plan, not verified provider model IDs or proof that either profile has been connected successfully.
What one story should look like
For a sample story—preventing a customer from submitting the same order twice—the proposed sequence is:
- The orchestrator selects an approved ticket and creates an isolated worktree.
- A fresh development session reads the ticket, acceptance criteria, relevant code and applicable lessons, then makes a candidate change.
- External checks inspect all changed files, run the relevant tests and assess whether the change stays within the approved scope.
- The candidate and test evidence are frozen for review. New files must be included, not just tracked-file diffs.
- A separate review session examines the candidate and evidence without access to the developer’s chat or self-assessment. It can approve, reject or return an unknown result; it should not rewrite the implementation.
- The workflow checks that the reviewed candidate has not changed, then commits through the orchestrator. Lessons from the task are saved only if supported by evidence.
- A later, fresh task can read the relevant lesson. One successful story would test the basic chain; a second story would be needed to show that experience actually carries forward.
The review stresses that ending an agent session is not the same as passing the story, and that a review session finishing is not the same as approving the change. If the candidate changes after approval, that approval must no longer apply.
Freeze V5; make V6 earn its launch
The recommendation is to mark V5’s execution effort as suspended or otherwise clearly discontinued, preserve its unfinished items and historical receipts, and create a separate plan and architecture decision for V6. It cautions against marking incomplete work as done or quietly treating old test results as new validation.
That freeze is a documentation and project decision; it does not prove that any running process has been stopped or that old launch paths have been physically disabled.
V6’s first engineering milestone should not be another home-grown controller. It should be proving the integration and a failure-closed review-and-commit path. The review’s proposed acceptance bar includes a real approved story, required tests, an external scope check, independent review and a native commit. It also calls for evidence that a new execution stage occurs after the local client disconnects, that rejection and service failures block commits, and that stopping a task actually ends its sessions and tool processes.
The central conclusion is less “replace everything” than “keep the analysis, simplify the runtime and demand evidence.” V5’s records and lessons may still be useful. But until Cline’s lifecycle, review isolation and the remote workflow are demonstrated, V6 is a direction—not a proven operating system.
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 "A leaner AI-agent stack starts by knowing what to stop building"?
The review recommends ending development on V5’s execution layer, while keeping its analysis methods, decision records and historical evidence.
What are the key points to validate first?
The review recommends ending development on V5’s execution layer, while keeping its analysis methods, decision records and historical evidence. Its proposed V6 stack has one orchestrator, one execution framework, external validation and a small, evidence based experience file.
What should I do next in practice?
Cline’s compatibility with the orchestrator, independent review isolation and reliable operation after a client disconnect remain unproven.