Four gates. Each runs after a phase, in every mode, at every depth. A gate that
fails is not a warning: the phase is redone.
<!-- maestro:runtime:node -->
<!-- maestro:delegation:native-explicit -->
This skill explicitly requests native subagent delegation for its independent
roles, including executors, audits, readers, review and diagnosis.
On Codex or dsh use the demand-loaded runtime recipe preflight and each dispatch
phase select. Never inherit coordinator history into an independent role.
Pass each dispatch its tier's model from
(
, or a contract-7 run with no
, passes none), effort as the session's: judges and repairs
, a build executor
its таск's tier, the existing-code reader, the commit reader and the probe
. A judge never runs
. Where the dispatch takes a
, open it with
— the brief's file name without
,
for the probe — then the таск's id when it serves one.
An independent reader, reviewer, or executor is a separate agent dispatch with
its own bounded inputs and a returned result. A summary you write yourself is
not that dispatch, even if it names the reader and its supposed findings. When
the host cannot provide the required separate context, keep the affected gate
pending and report the missing capability; do not record a pass.
For blind readers, start a fresh context containing only the allowed handoff;
do not fork the orchestrator's reasoning or earlier file reads into that
reader. Confirm that the separate agent actually returned before recording its
finding or pass.
| Gate | After phase | Pass condition |
|---|
| G1 | briefing | Every требование has a status and no unexplained open entry; contract 5 also requires a fresh returned independent source audit and frozen initial agreement |
| G2 | spec | Every live требование is in-spec, placeholder, deferred, or dropped with zero left open, and the independent intent reader and separate raw-reference reader return no unresolved mandatory gap, and the contextual architecture reader (when relevant decisions exist) finds no unresolved contradiction |
| G3 | plan | Every in-spec or placeholder требование maps to at least one таск, and every таск traces back to at least one требование; таски sharing a file are ordered, and the task and plan-consistency readers find no gap |
| G4 | acceptance | Current required checks and complete coverage pass after blind discovery against manifest/additions; failures give , missing or stale evidence gives ; contract 6 closes with no open defect |
At G4 the reader has
, the dated additions in
, and the
running build, and does
not have
, the plan, the task files, the
review notes, or the бриф's original text. The withholding is the mechanism: a
reader who has seen the specification confirms the specification, and a reader
holding the бриф's original text beside the манифест answers from the looser of
the two exactly where they disagree.
The original text is withheld and the additions are not: the text became
the манифест agreed at G1, and an addition never passed that gate.
When a gate fails, open
references/failure-modes.md
— the catalogue of
excuses and red flags. It is read at the failure and closed again; nothing in it is a rule, and keeping
it in context is how a catalogue turns into one.
A failed gate returns control to the phase it follows, which runs again with the gate's findings as input.
A gate may fail twice on the same finding; on the third failure the run stops and reports what cannot be
satisfied rather than looping. A stop that asks nothing writes
, never with
.
G1–G3 run at most 60 rounds, whatever each one found: a failed round appends one line, opening with its
number, to the gate's
; at the sixtieth the run stops on a question its phase file words.
A gate is never passed with notes — findings are acted on, or recorded as an explicit deferral
against a requirement id, which itself changes that requirement's status.
G4 is the exception to the first sentence. Приёмка re-run against the same
манифест and the same build finds the same thing, so a failed G4 does not send
its phase back. It sends sideways: the отчёт is written anyway — it is the record
of what disagreed — and each disagreement travels to Repair through the таски
carrying its
. Приёмка runs again once the build has changed, and appends
its round to the отчёт.
共有四个关卡。无论模式或深度如何,每个阶段结束后都会运行这些关卡。关卡失败不是警告,而是要求重做该阶段。
<!-- maestro:runtime:node -->
<!-- maestro:delegation:native-explicit -->
本 Skill 明确要求为独立角色使用原生子 Agent 委派,包括执行者、审计员、读取者、审查和诊断。
在 Codex 或 dsh 上,使用按需加载的运行时配方预检和每个派发阶段的选项。绝不能把协调器的历史上下文继承到独立角色中。
向每次派发传入其层级对应的模型,模型来自
(使用
;contract-7 运行没有
时则不传),并沿用会话的 effort:判定者和修复使用
,构建执行者使用其任务层级对应的模型,现有代码读取者、提交读取者和探针使用
。判定者绝不能使用
。当派发包含
时,以
打开它——其中
是去掉
的简报文件名,探针使用
——随后在适用时加上任务 ID。
独立读取者、审查者或执行者是一次独立的 Agent 派发,拥有各自有边界的输入并返回结果。你自己编写的总结不是该派发,即使其中写明了读取者及其预期发现。当宿主无法提供所需的独立上下文时,应让受影响的关卡保持 pending,并报告缺失的能力;绝不能记录为通过。
对于盲读,应从只包含允许交接内容的新上下文开始;不要把协调器的推理过程或先前读取的文件复制给读取者。记录发现或通过之前,必须确认独立 Agent 确实已经返回。
| 关卡 | 所在阶段之后 | 通过条件 |
|---|
| G1 | 简报澄清 | 每项需求都有状态,且不存在无法解释的开放条目;contract 5 还要求获得全新返回的独立来源审计,并冻结初始共识 |
| G2 | 规格 | 每项有效需求都处于规格内、占位、延期或已丢弃状态,开放项为零;并且独立意图读取器和独立的原始参考读取器均未发现尚未解决的强制性缺口,相关决策存在时,上下文架构读取器也未发现尚未解决的矛盾 |
| G3 | 计划 | 每项规格内或占位需求都至少映射到一个任务,并且每个任务都能追溯到至少一项需求;共享文件的任务具有明确顺序,任务读取器和计划一致性读取器均未发现缺口 |
| G4 | 验收 | 针对需求清单和新增内容进行盲发现后,当前必需检查和完整覆盖均通过;失败项标记为 ,证据缺失或过期标记为 ;contract 6 在没有开放缺陷时以 关闭 |
在 G4,读取器拥有
、
中带日期的新增内容和当前构建,但
没有 、计划、任务文件、审查记录或简报中的原始文本。这种限制本身就是机制:见过规格的读取器只会确认规格,而同时持有原始文本和需求清单的读取器,会在两者存在分歧之处依据约束更宽松的一方给出答案。
**原始文本会被保留,而新增内容不会。**原始文本已成为 G1 同意的需求清单,而新增内容从未通过那个关卡。
关卡失败时,打开
references/failure-modes.md
——这是借口和危险信号的目录。它只在失败时读取,随后再次关闭;其中没有任何内容是规则,而把它保留在上下文中会让一份目录变成另一份规则。
关卡失败后,控制权返回它所跟随的阶段,该阶段带着关卡发现作为输入重新运行。同一发现导致关卡两次失败后,第三次失败时运行应停止并报告无法满足的内容,而不是继续循环。一个不询问任何问题的停止会写入
,并且绝不携带
。无论每个关卡发现了什么,G1–G3 最多运行 60 轮:失败轮次会向该关卡的
追加一行,并以轮次编号开头;到第六十轮时,运行按照相应阶段文件中的措辞停止并提出问题。
关卡绝不能带着备注通过——发现要么已经处理,要么被记录为针对某个需求 ID 的明确延期,而延期本身会改变该需求的状态。
**G4 是第一句话的例外。**针对同一需求清单和同一构建重新执行验收,会再次发现同样的问题,因此失败的 G4 不会让验收阶段倒退。它会横向转移:报告仍会编写,因为报告需要记录发生分歧的内容;每个分歧都会通过携带其
的任务进入 Repair。构建发生变化后,验收会再次运行,并把新一轮结果追加到报告中。