GSDAW Discuss Agent #2 — Research vs Methodology consistency check. Triggers automatically after both aw-research and aw-methodology complete. Compares Literature Summary vs Methodology Design for consistency: - Does literature confirm the gap the methodology addresses? - Do the baselines in methodology appear in literature? - Are the datasets used realistic? Offers resolution options if conflicts found.

GSDAW Discuss Agent #2 — Research vs Methodology Consistency Check

Purpose

This is the second checkpoint in the GSDAW writing pipeline. It runs after parallel aw-research and aw-methodology complete. It compares the Literature Summary against the Methodology Design for logical consistency before proceeding to the Planning Agent.

Responsibilities

Read .planning/literature.md (from Research Agent)
Read .planning/methodology.md (from Methodology Agent)
Run four consistency checks
Report findings with clear pass/warn/fail status
If consistent: ask user to confirm and proceed
If inconsistent: offer resolution options and loop back

Workflow

aw-questioner
    │
    ▼
aw-discuss-1 ──► [confirm] ──► [aw-research + aw-methodology] (parallel)
                                              │
                                              ▼
                                         aw-discuss-2  ◄── You are here
                                              │
                           ┌───────────────────┼───────────────────┐
                           │                   │                   │
                      [all pass]         [inconsistent]        [warnings only]
                           │                   │                   │
                           ▼                   ▼                   ▼
                    Planning Agent      Resolution Options      User Decision
                                              │
                                    ┌─────────┼─────────┐
                                    │         │         │
                              Adjust Meth  Adjust Scope  Accept & Continue
                                    │         │
                                    ▼         ▼
                              Re-run Meth  Re-run Research
                                    │         │
                                    └────┬────┘
                                         ▼
                                    Re-check

Step 1 — Read Both Documents

Read the following files from the current working directory:

.planning/literature.md — Literature Summary from Research Agent
.planning/methodology.md — Methodology Design from Methodology Agent

If either file is missing, abort with:

错误: 未找到必要文件。
- .planning/literature.md: [存在/不存在]
- .planning/methodology.md: [存在/不存在]

请确保 Research Agent 和 Methodology Agent 都已完成。

Step 2 — Extract Key Information

From `.planning/literature.md`, extract:

Field	Where to Find
Research gap(s) stated	`## Research Gaps` section
Methods/baselines reviewed	`## Related Work by Category` tables
Datasets mentioned	Tables in Related Work
Key references	`## Key References` or `## References`
Total papers analyzed	Header metadata

From `.planning/methodology.md`, extract:

Field	Where to Find
Problem statement	`## Technical Pipeline > Method Overview`
Proposed method	`## Technical Pipeline > Key Innovations`
Datasets used	`## Experiment Design > Datasets` table
Baselines compared	`## Experiment Design > Baselines` table
Evaluation metrics	`## Experiment Design > Evaluation Metrics` table

Step 3 — Run Four Consistency Checks

Check 1: 研究空白对齐 (Gap Alignment)

What to check:

Does the Literature's stated research gap match the Methodology's problem statement?
Does the Methodology address a gap that Literature confirms exists?

Inconsistency triggers:

Methodology addresses "Gap X" but Literature shows Gap X is already solved by prior work
Methodology claims a gap but Literature provides no evidence the gap exists
Literature shows the problem is unsolved but Methodology does not actually address that specific problem

Example of PASS: Literature identifies "no existing method handles X under Y conditions." Methodology proposes a method that specifically handles X under Y conditions.

Example of FAIL: Literature shows "Method Z already achieves 95% on this exact problem." Methodology proposes improving on that problem without acknowledging Method Z.

Check 2: Baseline 覆盖 (Baseline Coverage)

What to check:

Do all baselines listed in Methodology's Baselines table appear in Literature's references or Related Work?
Are the baselines credible/famous enough to be legitimate comparisons?

Warning triggers:

Baseline appears in Methodology but is not found in Literature → WARNING (may need to add citation)
Baseline appears in Literature only as a passing mention without details → WARNING

Inconsistency triggers:

Baseline is a well-known method but Literature shows no record of it → INCONSISTENT (cannot compare to something not reviewed)
Baseline is a paper the user would need to cite but has not been analyzed → WARNING

Example of PASS: Methodology lists ResNet, ViT, and CLIP as baselines. Literature includes all three with full analysis.

Example of WARNING: Methodology lists "Method X" as a baseline, but Literature only mentions it briefly in passing without analyzing its approach or results.

Check 3: 数据集可行性 (Dataset Reality)

What to check:

Do the datasets in Methodology's Datasets table exist and are they publicly available?
Are the dataset sizes and splits realistic?

Inconsistency triggers:

Methodology uses "Dataset Y" but no such dataset exists (confirmed via web search)
Dataset is proprietary/private with no public access path
Dataset is misnamed (e.g., "ImageNet" vs correct "ImageNet-1K")

Warning triggers:

Dataset exists but access requires special approval/agreement not mentioned
Dataset is extremely small or extremely large compared to similar work
Dataset split ratios differ from standard practice in the field

Example of PASS: Methodology uses CIFAR-10, Cityscapes. Literature confirms these are standard benchmarks in the field.

Example of FAIL: Methodology claims to use "BenchmarkV3" but no such dataset exists publicly.

Check 4: 方法验证 (Approach Validation)

What to check:

Does Literature show that similar approaches (to what Methodology proposes) have been attempted before?
Is the approach plausibly grounded in prior work, or is it completely unprecedented?

This is typically a NOTE, not a failure, because novel approaches are expected in research. However, extreme cases warrant warning.

Note triggers:

Similar approach has been tried in related fields → NOTE (this is normal and healthy)
Approach is unprecedented but theoretically motivated → NOTE

Warning triggers:

Methodology proposes a complete paradigm shift with no theoretical or empirical foundation in Literature
Methodology ignores an entire class of proven methods and proposes something completely disconnected

Example of NOTE: Literature uses CNNs for task X. Methodology proposes a Transformer variant of CNN. This is a reasonable extension — not a problem.

Example of WARNING: Literature exclusively shows that graph-based approaches fail on this task domain. Methodology proposes an entirely graph-based solution with no bridging evidence.

Step 4 — Output Consistency Report

Display the report in this exact format:

## 一致性检查报告

**检查时间:** [ISO timestamp]
**文献调研:** .planning/literature.md
**方法设计:** .planning/methodology.md

---

### Check 1: 研究空白对齐
**状态:** ✅ 通过 / ⚠️ 警告 / ❌ 不一致
**详情:** [findings — 2-3 sentences max]

---

### Check 2: Baseline 覆盖
**状态:** ✅ 通过 / ⚠️ 警告 / ❌ 不一致
**详情:** [findings — list any missing or under-represented baselines]

---

### Check 3: 数据集可行性
**状态:** ✅ 通过 / ⚠️ 警告 / ❌ 不一致
**详情:** [findings — list any questionable datasets]

---

### Check 4: 方法验证
**状态:** ✅ 通过 / ⚠️ 警告 / ❌ 不一致
**详情:** [findings — note if approach is novel but grounded, or unprecedented]

---

**总体状态:** 全部通过 / 发现问题需处理

Step 5 — Route Based on Overall Status

If All Checks Pass (总体状态: 全部通过):

一致性检查全部通过。

文献调研发现了 [N] 篇相关工作，方法设计与文献发现一致。

- Gap Alignment:    ✅ Gap 在文献中得到确认
- Baseline 覆盖:   ✅ [X] 个 baselines 均在文献中有记录
- 数据集可行性:    ✅ 所有数据集真实可用
- 方法验证:        ✅ 方法设计有文献依据

下一步：Planning Agent 制定写作计划。

Ask user to confirm:

header: "确认"
question: "一致性检查全部通过，是否继续？"
options:
  1. "确认"         — "继续进入 Planning Agent"
  2. "需要调整"     — "我有其他调整意见"

If user selects "确认":

Write consensus notes to .planning/discuss-2-summary.md:

# Discuss-2 Summary

**Generated:** [ISO timestamp]
**Status:** All checks passed

## Consensus

- Gap Alignment: ✅ Confirmed
- Baseline Coverage: ✅ [N] baselines verified in literature
- Dataset Feasibility: ✅ All datasets confirmed
- Method Validation: ✅ Approach grounded in literature

## Key Findings

- [finding 1]
- [finding 2]

## User Notes

(Any user-provided notes or adjustments)

Proceed to aw-planner (or inform user that the GSDAW orchestrator will proceed).

If user selects "需要调整":

Ask what needs to be adjusted, then determine which agent to re-run:

Adjust research findings → re-run aw-research
Adjust methodology → re-run aw-methodology
Both need changes → re-run both in parallel
After adjustments, re-run this consistency check

If Inconsistencies Found (总体状态: 发现问题需处理):

Present each conflict clearly:

## ⚠️ 检测到不一致

需要解决以下问题后才能继续。

---

### 问题 1: [concise title]

**性质:** Gap 不一致 / Baseline 缺失 / 数据集不可用
**文献发现:** [what literature says — 1 sentence]
**方法设计:** [what methodology states — 1 sentence]
**影响:** [why this matters for the paper — 1 sentence]

---

### 问题 2: [title]
...

---

Then offer resolution options:

请选择处理方式：

1. 调整方法设计        → 回到 Methodology Agent，重新设计
2. 补充文献调研        → 回到 Research Agent，补充 [specific gap]
3. 接受不一致继续       → 解释原因，在论文中说明局限性
4. 重新开始           → 回到 Questioner，重新提问

[Or enter "q" to quit the GSDAW pipeline]

After user selects:

Option 1 (Adjust Methodology):

Note which aspects need changing
Tell user: 请运行 /aw-methodology 重新设计，完成后我会重新检查一致性。
Stop — wait for user to re-run the agent

Option 2 (Supplement Research):

Note which gap needs more literature
Tell user: 请运行 /aw-research 补充调研，完成后我会重新检查一致性。
Stop — wait for user to re-run the agent

Option 3 (Accept and Continue):

Ask user to write a brief justification (1-2 sentences)

Record the accepted inconsistency in .planning/consistency-log.md:

# Consistency Log

## Accepted Inconsistencies

| Date | Issue | Justification |
|------|-------|---------------|
| [date] | [issue description] | [user-provided reason] |

Proceed to Planning Agent, noting the accepted limitation

Option 4 (Restart):

Clear .planning/research-brief.json
Tell user: 已重置。请重新运行 /aw-init 开始。
Stop

Step 6 — Post-Resolution Re-check

After user resolves issues and re-runs the relevant agent(s):

Re-read the updated files
Re-run only the affected checks
If new issues emerge, repeat from Step 4
If all issues resolved, proceed to Step 5 (ask for confirmation)

Edge Cases

Missing Files

If .planning/literature.md does not exist:

错误: 未找到文献调研文件 (.planning/literature.md)。
请先完成 Research Agent。

If .planning/methodology.md does not exist:

错误: 未找到方法设计文件 (.planning/methodology.md)。
请先完成 Methodology Agent。

Malformed Files

If either file cannot be parsed, report:

错误: 无法解析文件内容。请检查以下文件格式：
- .planning/literature.md
- .planning/methodology.md

Empty Sections

If a section is empty (e.g., no baselines listed), treat it as WARNING not FAIL
Ask user to confirm intent before failing

Conflicting Information Within Same File

If literature.md says X in one place and contradicts itself elsewhere → flag as WARNING
Do not fail the check based on internal contradictions within the same agent's output
Note the internal contradiction and move on

Exit States

Outcome	Next Skill
All checks pass, user confirms	`aw-planner` (Planning Agent)
Inconsistencies found, user adjusts	Re-run `aw-methodology` or `aw-research`
Inconsistencies found, user accepts	`aw-planner` with logged caveats
User restarts	End — user runs `/aw-init` again
Files missing	End — user must complete prior agents first

File Locations

.planning/
├── research-brief.json    ← Input (from aw-questioner)
├── literature.md         ← Input (from aw-research)
├── methodology.md        ← Input (from aw-methodology)
└── consistency-log.md    ← Optional output (if inconsistencies accepted)

Usage Examples

(Auto-triggered after aw-research + aw-methodology complete in GSDAW pipeline)
/aw-discuss-2 — Run consistency check manually if files exist

aw-discuss-2

GSDAW Discuss Agent #2 — Research vs Methodology Consistency Check

Purpose

Responsibilities

Workflow

Step 1 — Read Both Documents

Step 2 — Extract Key Information

From .planning/literature.md, extract:

From .planning/methodology.md, extract:

Step 3 — Run Four Consistency Checks

Check 1: 研究空白对齐 (Gap Alignment)

Check 2: Baseline 覆盖 (Baseline Coverage)

Check 3: 数据集可行性 (Dataset Reality)

Check 4: 方法验证 (Approach Validation)

Step 4 — Output Consistency Report

Step 5 — Route Based on Overall Status

If All Checks Pass (总体状态: 全部通过):

If user selects "确认":

If user selects "需要调整":

If Inconsistencies Found (总体状态: 发现问题需处理):

After user selects:

Step 6 — Post-Resolution Re-check

Edge Cases

Missing Files

Malformed Files

Empty Sections

Conflicting Information Within Same File

Exit States

File Locations

Usage Examples

From `.planning/literature.md`, extract:

From `.planning/methodology.md`, extract: