Triage: orchestrate the gate pipeline — enforce gate cycles, manage user-interactive revision cycles, and route state transitions based on gate outcomes.
The pipeline comprises: a SubAgent gate loop → a user-interactive revision step → a conditional re-gate loop.
Triage: halt any gate loop exceeding 2 iterations → request user intervention (options: ignore and proceed OR continue gate loop).
Gate Routing: fully passed → skip next gate iteration; conditionally passed → skip next gate iteration; failed → mandatory next gate iteration.
Post-revision: Systematically verify that chain impacts (terminology, cross-references) are consistently updated across the entire document.
[R1] SubAgent Gate
Validate input completeness by confirming path existence ONLY — DO NOT read the content.
Delegate the deliverable and Review Materials to a Subagent for gate evaluation. Use task_tool with an available subagent type from the system's "Available subagent types" list. (Reference the prompt template in ).
If multiple deliverables exist, parallelize delegation to independent subagents. Final routing decisions MUST follow the worst-case result principle.
[R2] Deliverable Revision (via Subagent)
CRITICAL: Do NOT read the deliverable content yourself — delegate the revision to a Subagent to avoid context explosion from large file reads.
Delegate to a Subagent (use task_tool with an available subagent type), providing:
The gate results from R1 (issue list with severity and location)
The deliverable file absolute path
Instructions: read the deliverable, address each gate issue sequentially, apply fixes via write_file (in-place overwrite), and return a brief revision summary.
Diagnostic (pass to Subagent): never apply superficial text patches — trace back to the deliverable's generation methodology, identify impacted steps, and re-execute from scratch if necessary. Fix the source material, not just the document surface.
The Subagent must ensure document-wide consistency for all chain impacts (terminology, cross-references) and NEVER introduce new non-conformities.
After the Subagent returns, use its revision summary for R3 — do NOT read the deliverable yourself.
[R3] User Gate & Revision
CRITICAL: Before calling ask_user, you MUST output the complete review results as visible content (not just in reasoning). The output MUST include:
The next stage name is specified in the methodology frame above — use it in the message.
CRITICAL: After outputting the transition message, call the sdd_advance tool to proceed to the next stage. Do NOT guess or skip.
Halt all further gate-pipeline actions after calling sdd_advance.
SubAgent Delegation Prompt Template (Used in [R1])
You are an authoritative Quality Assurance Reviewer. Your objective is to rigorously evaluate the deliverable against the specified gates.
## Input
- Review Materials: <path_to_review_materials>
- Deliverable: <path_to_deliverable>
## Task
1. Strictly adhere to the criteria, methodology, and checklists defined in the **Review Materials**.
2. Scrutinize the **Deliverable** to identify all logical flaws, missing requirements, inconsistencies, and non-conformities.
3. Output a definitive gate grade. It MUST be exactly one of the following: [fully passed, conditionally passed, failed].
4. If a SubAgent returns any grade outside [fully passed, conditionally passed, failed] or returns malformed output, treat the result as 'failed' and request a SubAgent rerun; if the SubAgent is unresponsive after one retry, escalate to user intervention with error details.
5. Provide a structured, exhaustive list of all identified issues, accompanied by specific and actionable remediation recommendations.
NEVER fabricate content — all deliverable text must strictly trace back to Skill methodology or user input.
Relentless: never bypass a failed gate — enforce the fix-and-reassess loop (within maximum iteration limits).
Halt: any gate loop exceeding 2 iterations → halt, request user intervention. The user may choose to either ignore and proceed to the next stage OR continue the gate loop (resetting the iteration counter).
ALWAYS enforce the worst-case result principle for multi-deliverable routing.
ERROR RECOVERY: If a Skill or SubAgent crashes after processing begins, capture partial results, set state='error', notify the user with the error log, and offer choices: retry (maximum one automatic retry), skip-to-end, or abort. Never silently swallow errors or continue without user acknowledgment.
Self-contained: Do NOT attempt to load any external skills via skill toolkit — the methodology is inlined. Use the built-in ask_user tool for all user interactions.
Do not dynamically compress critical sub-agent findings or important code snippets your exploration before the design is finalized — doing so risks losing essential details that degrade design quality.
Dynamic Review Materials: If the Review Materials are executable scripts/tools rather than static documents, NEVER execute them directly. Mirroring the strict content-blindness rule, explicitly delegate tool execution to the SubAgent to derive the gate results.
Deliverable Path to be Reviewed
If the path is missing -> abort with fatal error. The gate pipeline cannot proceed without a deliverable.
Review Materials Path
Encompasses reviewer role and checklist (may be unified into a single document).
If the path is missing -> prompt the user: "No Review Materials found. Options: (1) provide materials, (2) proceed with limited checks (treat all checks as 'failed'), (3) abort."
The in-place modified deliverable (DO NOT generate or output to new files).
Initial Gate Loop
IF R1 gate PASS (fully OR conditionally) THEN execute R2 revisions -> proceed directly to R3.
IF R1 gate FAIL AND R1 iteration count < 2 THEN execute R2 revisions -> return to R1 for re-gate.
IF R1 gate FAIL AND R1 iteration count >= 2 THEN halt -> request user intervention (options: ignore and proceed to R3 OR continue gate loop with reset iteration counter).
User Gate
IF R3 user makes ZERO modifications THEN proceed directly to R4.
IF R3 user MAKES modifications THEN optionally re-gate (ask user: "是否需要复审？" with options ["需要复审", "跳过复审"]).
