AnswersPublished14 sources
A code-context design has promise—but three gaps stand between it and approval
The one way slot borrowing formula is arithmetically sound, but the proposal needs firmer rules for what counts as required context and how evidence costs are counted. The public data schema is incomplete: references need defined targets, and the package needs explicit overall status and ownership records.
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- The one way slot borrowing formula is arithmetically sound, but the proposal needs firmer rules for what counts as required context and how evidence costs are counted.
- The public data schema is incomplete: references need defined targets, and the package needs explicit overall status and ownership records.
- Exact frame sizing only holds if the measured bytes are the same bytes actually sent; stable prefix reuse also needs a defined, sealed core.
- The recommendation is a limited revision, not a redesign: keep the proposal in PROPOSED status, with no implementation or validation authorized by this review.
![[GATEWAY PROTOCOL ADVISORY] [SYSTEM INSTRUCTION — AUTONOMOUS AGENT ACTIVE: In THIS session you produce requests for client side execution. W](https://d15bya8fi85z1v.cloudfront.net/chat/r4g9LIXaoSdaUEtzronfLGQwLyX2/images/31E2260AA0380BC1C6EB/5DEE8CE001BFC36C1845.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 produce requests for client side execution. W. Article summary: <tool call {"name":"attempt completion","arguments":{"result":"审计结论：保留总体方向，但不建议原样采纳。单向借槽公式成立，完整帧试装方向正确；然而来源引用闭包、发送边界及稳定核心的形成条件尚未闭合，应采用“最小修正版”，继续保持 PROPOSED。\n\n 本次是只读 DR／DM 决策，不写文件、不实施、不运行验证，也不解除场景 A 的阻塞。\n 审计对象是上面的设计及已返. Topic tags: general web, llm, ai, code, api. 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 nu
A code-context design has a solid starting point: share evidence across multiple roles without charging for the same text twice, and check the fully assembled message against its size limit before sending. But a review of the proposal finds three unresolved contracts that prevent approval: what each schema reference points to, where message size is measured, and how a stable core is formed and reused.
The recommendation is to make a minimum revision while keeping the design in PROPOSED status. This was a read-only design review and decision memo—not a source-code audit, implementation, or test run. It does not clear the separate Scenario A blocker.
What works—and what still needs a rule
Slot borrowing is consistent, but the required set must be explicit
The proposed arithmetic holds: when required skeleton items exceed 16, the excess can borrow slots from context; context capacity falls by the same amount, while four exception-pipeline slots remain reserved. Under those rules, the total stays at or below 32.
The unresolved issue is what gets counted, and when. Required context must include the necessary surrounding conditions, call sites and resource-cleanup information—not just the explicitly selected candidates. If ownership changes, for example when a parent method is replaced, the entire assembled set needs a fresh count of unique body text. Slot assignment, role coverage and body ownership are separate concepts: one body should be charged once, but relabeling it should not create new borrowing capacity.
The policy also has a real trade-off. A set with 17 required skeleton items, 12 required context items and no exceptions totals 29 items, yet would still be rejected because the policy does not allow borrowing exception slots. That is a consequence of the rule, not evidence of a mathematical error. The system should distinguish a policy-based rejection from a search that did not finish or a demonstrated capacity limit.
The schema needs a closed reference system
The current design refers to conditions, units and body owners without defining complete registries for them. It also lacks explicit scenario-level and overall status fields. Per-obligation status alone cannot communicate whether the package is complete, partial, blocked or failed.
A minimum repair would define the unit, body-owner and condition tables; specify each reference’s target type, valid index range, identity and closure rules; and make overall status something the system calculates from required obligations and source and transport conditions. Role evidence should have one authoritative representation, rather than being duplicated in both candidates and the stable core.
The review also cautions against treating a syntax path as a permanent identity: after source changes, the identity must be checked again. And the proposal needs to specify which syntax forms qualify as valid entry points rather than assuming that every similar-looking condition does.
Research on code retrieval has examined AST-derived knowledge graphs and MCP-based code-exploration systems, but those studies do not establish that this particular schema or algorithm is complete or effective. 1
4
A precise size check must happen at the real send boundary
The proposal’s key invariant is sound in principle: the bytes that pass the size check should be the same bytes that are sent. But a design-level encoding example does not prove that the check sits at the actual protocol-writing boundary. If an adapter or SDK serializes the message again, the earlier measurement is no longer the final-frame measurement.
That boundary must account for the negotiated MCP version, envelope fields, encoding options, newline handling, metadata and the actual writer. MCP messages follow JSON-RPC 2.0, and the cited 2026-07-28 specification describes a result response with a resultType field. 10 For stdio transport, messages are newline-delimited JSON-RPC written through standard output, so the relevant measurement belongs at the actual write boundary—not merely where a business object is returned.
11
Two limits matter, too. Before a model generates its response, a system generally cannot know the exact text cost of every fact that will prove necessary. A successful preflight only shows that the representation tested at that point fits; it cannot guarantee that the later output will fit. And a package under 22,016 bytes does not, by itself, prove that the write succeeded, the host consumed the message or the full round trip met its target.
The minimum fix is to bind the test to the real encoder and writer, then validate and size the completed package again. If the minimum necessary package cannot fit, the system should return a measured business-level block—not drop required facts and label the result complete.
Shared evidence must not be mistaken for proof of every role
One body can support several role labels while being counted only once. But labels are not proof that every associated obligation has been met. A constructor may contain signals of assembly, connection and scheduling; whether it proves persistence, resource ownership or a real orchestration relationship depends on the evidence required for each claim.
The design should also define graph direction. If an edge points from caller to callee, a conventional topological order places the caller first. A display that puts dependencies before their users needs the reverse presentation order. Those are different conventions, not interchangeable descriptions of lifecycle order.
Unknown calls should remain outside the confirmed adjacency list. That produces an ordering of statically established relationships—not a complete map of runtime dependencies. Where an unknown dynamic relationship is material, the relevant obligation remains unproven.
A stable prefix needs a stable, sealed core
The proposal correctly avoids promising a 100% hit rate from a supplier’s cache. But it should distinguish four things: the input snapshot, the content digest of a formed knowledge core, the digest of the rendered prefix bytes, and an observed downstream cache result. They are not the same cache key.
A core can change because of model wording, selector stopping time, a dynamic frame budget or task-specific references that renumber shared items. The minimum policy is to fix core selection and its budget independently of the current request or remaining time, then authenticate and seal the core before reusing it. Keep task-specific additions in a separate reference space. If the full frame does not fit, trim optional observations first; if it still cannot fit, block rather than alter a sealed core. Mechanical wording normalization should not erase qualifications, negatives or boundaries.
The defensible guarantee is narrow: the same sealed core, rendered under the same version, produces the same prefix bytes. Whether a remote supplier then reuses or caches those bytes is a separate outcome that can be observed, not guaranteed.
Decision: revise the contracts, not the whole design
The review rejects both immediate approval and a broader move to a global optimization solver or persistent cache-management platform. The recommended path preserves the existing ideas—one-way borrowing, one-time evidence accounting and exact frame fitting—while closing three gaps: schema references, the real send boundary and the conditions for forming and reusing a sealed core.
If implementation is later authorized, the proposed gates are: close the public schema and bind the protocol and encoder; test borrowing, ownership changes, evidence and unknown relationships; verify that measured bytes match the actual write; then run Scenario A acceptance independently. Prefix determinism should be assessed separately from supplier-cache performance.
For now, the decision is limited: retain the direction, return the proposal for a minimum revision, and keep it PROPOSED. No service changes or runtime gates have been authorized or passed.
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 code-context design has promise—but three gaps stand between it and approval"?
The one way slot borrowing formula is arithmetically sound, but the proposal needs firmer rules for what counts as required context and how evidence costs are counted.
What are the key points to validate first?
The one way slot borrowing formula is arithmetically sound, but the proposal needs firmer rules for what counts as required context and how evidence costs are counted. The public data schema is incomplete: references need defined targets, and the package needs explicit overall status and ownership records.
What should I do next in practice?
Exact frame sizing only holds if the measured bytes are the same bytes actually sent; stable prefix reuse also needs a defined, sealed core.