Imported from nolood/planpipe (skills/task-synthesis/SKILL.md). Install upstream with npx skills add nolood/planpipe --skill task-synthesis. Copyright stays with the author.

Task Synthesis & Agreement — Stage 3

You are executing Stage 3 of the planning pipeline. Your job is to synthesize the multi-stream analysis from Stage 2 into a single coherent task model and resolve contradictions.

You do NOT write plans, design solutions, or write code. You synthesize — nothing more. Stage 2 investigated. Stage 3 consolidates. Stage 4 will design the solution AND present the combined review (understanding + design) to the user.

Think of it this way: Stage 2 produced three independent views of the same task. They overlap, they might contradict, they each emphasize different things. Your job is to merge them into one honest picture. User confirmation happens later — jointly with the design review in Stage 4 — to reduce review fatigue and let the user evaluate understanding and design as a coherent whole.

Input Requirements

This stage requires Stage 2 output.

Before doing anything else, verify you have one of these input sets:

Preferred (new Stage 2 format):

• stage-2-handoff.md — self-contained handoff document with consolidated findings from all three analyses. This is the single entry point.

Full context (recommended to also load):

• product-analysis.md — detailed product/business analysis
• system-analysis.md — detailed codebase/system analysis
• constraints-risks-analysis.md — detailed constraints/risks analysis

Also check for:

• stage-1-handoff.md or requirements.draft.md — original task statement for reference
• clarifications.md — any prior Q&A that shaped the understanding
• Memobank / memory directory — search for related context, past decisions, known patterns

If stage-2-handoff.md is missing or doesn't reference completed analyses, stop and tell the user — send it back to Stage 2.

Process

The stage runs as: synthesize → critique → refine → package for combined review in Stage 4. User confirmation does NOT happen here — it happens jointly with the design review in Stage 4.

Step 1: Load and Cross-Reference All Inputs

Read all Stage 2 artifacts. Build an internal picture:

1. Read stage-2-handoff.md for the consolidated view
2. Read each detailed analysis for the full depth
3. Cross-reference: where do analyses agree? Where do they diverge? Where does one stream flag something the others missed?

Create a working list of:

• Agreements — things all three analyses align on
• Tensions — places where analyses say different things or emphasize conflicting priorities
• Gaps — things that none of the analyses covered well enough
• Open questions — unresolved from Stage 2

Step 2: Synthesize Unified Task Model

Merge everything into a single structure. This is not copy-pasting sections together — it's actual synthesis: resolving overlaps, choosing between contradicting views (with reasoning), and filling in the gaps.

Build analysis.md using the template from the Artifact Templates section below.

The synthesis covers:

• Task goal — one clear formulation drawn from product analysis + original requirements
• Key scenarios — main flow and mandatory edge cases, cross-checked against system change points
• System scope — which modules/areas are involved, validated against the system analysis
• Constraints — merged from all sources, deduplicated, prioritized
• Risks — consolidated, recalibrated where analyses disagreed on likelihood/impact
• Candidate solution directions — not solutions, but general directions (minimal vs. systematic, refactor vs. extend, etc.) informed by what the analyses revealed

When resolving contradictions:

• State what each source says
• Explain which view you chose and why
• If you can't resolve it — mark it as an explicit open question for the user

Step 3: Critique the Synthesis

Spawn a Synthesis Critic subagent to independently review the unified model.

1. Use the Synthesis Critic definition from the Agent Definitions section below
2. Use the Agent tool with:
   • name: "synthesis-critic"
   • subagent_type: "general-purpose"
   • prompt: the FULL content of the <synthesis-critic> definition combined with the input data below — the agent definition IS the prompt, do not summarize or skip it
3. Input data to append to the prompt: the synthesized analysis.md content + all Stage 2 analyses + Stage 1 content

The critic checks:

• Internal consistency of the synthesized model
• Whether contradictions from analyses were actually resolved (not just hidden)
• Whether assumptions were promoted to facts without evidence
• Whether the model accurately reflects what the analyses found
• Whether anything important was lost during synthesis

Save the critic's feedback. If the critic finds significant problems, fix them before presenting to the user.

Step 4: Build Agreement Package

Once the synthesis passes critique (or has been refined), split the model into agreement blocks — discrete, reviewable chunks for the user.

Build agreement-package.md using the template from the Artifact Templates section below.

The blocks are:

Block 1 — Goal & Problem Understanding

• Did we understand the task correctly?
• Are we solving the right problem?
• Is the expected outcome what the user needs?

Block 2 — Scope

• What's included?
• What's excluded?
• Was anything added that shouldn't be there?

Block 3 — Key Scenarios

• Which scenario is primary?
• Which edge cases are mandatory?
• Which scenarios can be deferred?

Block 4 — Constraints

• Can the API change?
• Is backward compatibility required?
• Are there technical/business/process constraints?
• Are there deadlines?

Block 5 — Candidate Solution Directions

• Which direction should planning take?
• Preference: minimal, safe, or systematic?

Each block should be:

• Short enough to review in 30 seconds
• Oriented toward confirmation or correction — not education
• Ending with a clear question: "Is this correct? What would you change?"

Step 5: Package for Combined Review

Do NOT present blocks to the user here. User review happens in Stage 4, jointly with the design review.

Build agreement-package.md using the template from the Artifact Templates section. This package will be included in Stage 4's combined review.

Build a draft agreed-task-model.md using the template from the Artifact Templates section. Mark it as status: draft — pending user confirmation in Stage 4. This draft reflects the best synthesis validated by the critic, but has not yet been confirmed by the user.

Why no user review here: Presenting understanding separately from design creates review fatigue (8-12 review points before any code). The user thinks holistically — they can't confirm Scope without thinking about Scenarios, and can't confirm Solution Direction without seeing the actual design. Combining these reviews in Stage 4 gives the user one serious review of "here's what we understood AND here's what we'll build" instead of two fragmented ones.

Step 6: Build Handoff

Once the draft task model and agreement package are ready, build the handoff document.

stage-3-handoff.md is the single entry point for Stage 4 (solution design). It packages the synthesized model with all context the design stage needs. Stage 4 should be able to read this file alone and have everything it needs to begin design AND to present the combined review to the user.

Important: The handoff must clearly indicate that the task model is a draft pending user confirmation. Stage 4 is responsible for presenting the combined review (understanding + design) and finalizing the model based on user feedback.

Save stage-3-handoff.md and tell the user that Stage 3 is complete — synthesis is done, and the combined review will happen after design in Stage 4.

Step 7: Report to the User

Present a brief summary:

• What was synthesized
• Key findings and resolved contradictions
• What the draft model contains
• Note: user review will happen jointly with design review in Stage 4

Then offer the user two options for continuing to Stage 4:

Option 1 — Continue in this session:

"Запустить Stage 4 (Solution Design) прямо сейчас в этой сессии? Там же будет combined review."

If the user agrees, invoke the /solution-design skill.

Option 2 — Continue in a new session: Provide a ready-to-paste block with actual paths filled in:

Запусти /solution-design

Task ID: {task-id}
Артефакты: .planpipe/{task-id}/ (stage-1/, stage-2/, stage-3/)

Artifact Templates

This stage produces up to four files. Every artifact must follow its template exactly. These templates are not optional — they ensure consistency across tasks and enable Stage 4 to parse the output reliably.

1. analysis.md

When: Always created. The synthesized analytical model of the task — before user agreement.

# Synthesized Task Analysis

## Task Goal
[One clear formulation of what needs to be achieved — synthesized from product analysis and original requirements]

## Problem Statement
[Why this task exists — the deeper motivation, synthesized from business intent and original problem statement]

## Key Scenarios

### Primary Scenario
1. [Trigger: what starts the flow]
2. [Step: what happens]
3. [Step: ...]
4. [End state: what the actor sees/has when done]

### Mandatory Edge Cases
- **[Edge case]:** [What happens and why it must be handled in this task]
- **[Edge case]:** [...]

### Deferred Scenarios
- **[Scenario]:** [Why it can be deferred — what the risk of deferring is]

## System Scope

### Affected Modules
| Module | Path | Role in Task | Change Scope |
|--------|------|-------------|-------------|
| [name] | `path/to/module` | [what changes and why] | small/medium/large |

### Key Change Points
| Location | What Changes | Why |
|----------|-------------|-----|
| `path/to/file:Function` | [what] | [why this change is needed] |

### Dependencies
- **[Dependency]:** [What it is, how it constrains the task]

## Constraints
- **[Constraint]:** [What it is, source, impact on planning]
- **[Constraint]:** [...]

## Risks

| Risk | Likelihood | Impact | Mitigation Direction |
|------|-----------|--------|---------------------|
| [risk] | low/medium/high | low/medium/high | [brief idea] |

## Candidate Solution Directions
- **[Direction name]:** [What it means, when it's appropriate, trade-offs]
- **[Direction name]:** [...]

## Resolved Contradictions
[Where analyses disagreed and how it was resolved]
- **[Topic]:** [Analysis A said X, Analysis B said Y. Resolution: Z because...]

## Remaining Open Questions
[Things that could not be resolved from analysis alone — need user input or will surface during planning]
- [Question 1]
- [Question 2]

## Critique Review
[Summary of synthesis critic's findings. What was flagged. What was fixed.]

2. agreement-package.md

When: Always created. The structured package for block-by-block user review.

# Agreement Package

> Task: [one-line task summary]
> Based on: Stage 2 analyses (product, system, constraints/risks)
> Purpose: Confirm or correct the synthesized understanding before planning

---

## Block 1 — Goal & Problem Understanding

**Our understanding:**
[2-3 sentences: what the task achieves and why it matters]

**Expected outcome:**
[What "done" looks like]

**Confirm:** Is this the right goal? Are we solving the right problem? Is this the outcome you need?

---

## Block 2 — Scope

**Included:**
- [What's in scope — specific items]

**Excluded:**
- [What's explicitly out of scope]

**Confirm:** Is the scope correct? Anything missing? Anything that shouldn't be here?

---

## Block 3 — Key Scenarios

**Primary scenario:**
[Brief description of the main flow]

**Mandatory edge cases:**
- [Edge case 1]
- [Edge case 2]

**Deferred (not in this task):**
- [Scenario that can wait]

**Confirm:** Is the primary scenario correct? Are the mandatory edge cases right? Can the deferred items really wait?

---

## Block 4 — Constraints

- [Constraint 1]
- [Constraint 2]
- [Constraint 3]

**Confirm:** Are these constraints accurate? Are there constraints we missed? Can any of these be relaxed?

---

## Block 5 — Candidate Solution Directions

Based on the analysis, we see these possible directions:

- **[Direction A]:** [Brief description — trade-offs]
- **[Direction B]:** [Brief description — trade-offs]

**Confirm:** Which direction do you prefer? Minimal and safe, or systematic and thorough? Any direction we should avoid?

3. agreed-task-model.md

When: Created as a draft during Stage 3 (pending user confirmation in Stage 4's combined review). Finalized by Stage 4 after user confirms.

# Agreed Task Model

> Status: [draft — pending user confirmation / confirmed]
> Agreed on: [date — filled when confirmed in Stage 4]
> Based on: Stage 2 analyses + synthesis critique

## Task Goal
[Final, user-confirmed goal statement]

## Problem Statement
[Final, user-confirmed problem description]

## Scope

### Included
- [Confirmed scope item 1]
- [Confirmed scope item 2]

### Excluded
- [Confirmed exclusion 1]
- [Confirmed exclusion 2]

## Key Scenarios

### Primary Scenario
1. [Step 1]
2. [Step 2]
3. [...]

### Mandatory Edge Cases
- **[Edge case]:** [Description]

### Explicitly Deferred
- **[Scenario]:** [Why deferred, user confirmed]

## System Scope

### Affected Modules
| Module | Path | Role in Task | Change Scope |
|--------|------|-------------|-------------|
| [name] | `path/to/module` | [role] | small/medium/large |

### Key Change Points
| Location | What Changes | Why |
|----------|-------------|-----|
| `path/to/file:Function` | [what] | [why] |

### Dependencies
- **[Dependency]:** [Impact on task]

## Confirmed Constraints
- **[Constraint]:** [Description — confirmed by user]

## Risks to Mitigate

| Risk | Likelihood | Impact | Mitigation Direction |
|------|-----------|--------|---------------------|
| [risk] | low/medium/high | low/medium/high | [direction] |

## Solution Direction
[User's confirmed preference — which direction, why, trade-offs accepted]

## Accepted Assumptions
[Assumptions the user explicitly accepted as safe to proceed with]
- [Assumption 1: why it's accepted]

## Deferred Decisions
[Decisions that were explicitly pushed to the planning stage]
- [Decision 1: why it's deferred]

## User Corrections Log
[What the user changed from the original synthesis — preserves the decision trail]
- **[Block/Topic]:** [What was proposed → What the user said → How the model was updated]

## Acceptance Criteria
[How to know the task is done correctly — derived from goal + user confirmations]
- [Criterion 1]
- [Criterion 2]

4. stage-3-handoff.md

When: Created when Stage 3 synthesis is complete — critic has validated, artifacts are ready. This is the primary input for Stage 4. User confirmation has NOT happened yet — it happens in Stage 4's combined review.

# Stage 3 Handoff — Task Synthesis Complete

> Status: draft — pending user confirmation in Stage 4

## Task Summary
[Synthesized task statement — 2-3 sentences a new team member could read and immediately understand]

## Classification
- **Type:** [feature / bug / refactor / integration / research / other]
- **Complexity:** [low / medium / high]
- **Primary risk area:** [technical / integration / scope / knowledge]
- **Solution direction:** [minimal / safe / systematic — synthesized from analyses]

## Synthesized Goal
[One clear sentence — from synthesis, pending user confirmation]

## Synthesized Problem Statement
[Why this matters — from synthesis, pending user confirmation]

## Synthesized Scope

### Included
- [Item 1]
- [Item 2]

### Excluded
- [Item 1]
- [Item 2]

## Key Scenarios for Planning

### Primary Scenario
1. [Step 1]
2. [Step 2]
3. [...]

### Mandatory Edge Cases
- [Edge case 1]
- [Edge case 2]

## System Map for Planning

### Modules to Change
| Module | Path | What Changes | Scope |
|--------|------|-------------|-------|
| [name] | `path/to/module` | [what and why] | small/medium/large |

### Key Change Points
| Location | What Changes | Why |
|----------|-------------|-----|
| `path/to/file:Function` | [what] | [why] |

### Critical Dependencies
- **[Dependency]:** [What it is, how it constrains planning]

## Constraints for Planning
- [Constraint 1: what and why — from analyses]
- [Constraint 2: ...]

## Risks to Mitigate

| Risk | Likelihood | Impact | Mitigation Direction |
|------|-----------|--------|---------------------|
| [risk] | low/medium/high | low/medium/high | [direction from analyses] |

## Product Requirements for Planning
- **Primary scenario:** [1-sentence summary]
- **Success signals:** [what to measure]
- **Minimum viable outcome:** [smallest valuable delivery]
- **Backward compatibility:** [what must not break]

## Solution Direction
[Synthesized approach — minimal/safe/systematic, with rationale from analyses. Pending user confirmation in Stage 4.]

## Assumptions (pending confirmation)
- [Assumption: why it seems safe — to be confirmed by user in Stage 4]

## Deferred Items
- [Item: what and why it's deferred]

## Acceptance Criteria
- [Criterion 1]
- [Criterion 2]

## Detailed References
[These files contain the full analysis and draft model:]
- `analysis.md` — synthesized task analysis
- `agreement-package.md` — agreement blocks for Stage 4's combined review
- `agreed-task-model.md` — draft task model (pending user confirmation)
- `product-analysis.md` — detailed product/business analysis (Stage 2)
- `system-analysis.md` — detailed codebase/system analysis (Stage 2)
- `constraints-risks-analysis.md` — detailed constraints/risks analysis (Stage 2)

Artifact Summary

Save all artifacts to .planpipe/{task-id}/stage-3/.

Done Criteria

Stage 3 is complete when all of these hold:

• Findings from Stage 2 analyses are synthesized into a unified model
• Contradictions between analyses are resolved or explicitly flagged
• Agreement package has been built (for use in Stage 4's combined review)
• Draft agreed-task-model.md has been created (pending user confirmation in Stage 4)
• Synthesis critic has confirmed the model is consistent enough for design
• stage-3-handoff.md has been created

Note: User confirmation does NOT happen in Stage 3. It happens in Stage 4's combined review (understanding + design together).

Failure Criteria

Stage 3 is NOT complete if any of these hold:

• Results from different Stage 2 analyses still conflict without resolution
• Assumptions were promoted to facts without evidence
• Synthesis critic found significant problems that were not addressed
• Blocking contradictions remain that would affect design
• Draft agreed-task-model.md has not been created
• agreement-package.md has not been created (needed for Stage 4's combined review)
• stage-3-handoff.md has not been created

Notes

• Synthesis, not planning. If you catch yourself sequencing work or designing architecture, stop — that's Stage 4. Your job is to consolidate understanding.
• No user review here. User review happens in Stage 4, jointly with the design. This reduces review fatigue — two serious reviews (after design, after decomposition) instead of five fragmented ones.
• Don't hide contradictions. If two analyses say different things, show both views and explain your resolution. Hiding a contradiction is worse than surfacing an uncomfortable truth.
• The critic catches your blind spots. You wrote the synthesis — you can't objectively review it. The critic can. When it flags something, investigate before dismissing.
• The draft model is the product. Everything else is working material. Draft agreed-task-model.md is what Stage 4 builds the design on and then presents to the user for combined confirmation.
• Memobank check. If the project has a memobank or knowledge store, check it for relevant context — past decisions on similar tasks, known patterns, prior agreements. Opportunistic — skip if nothing exists.
• Templates are not optional. Stage 4 depends on consistent structure. An artifact that doesn't follow the template is incomplete.
• Subagent prompts = agent definitions. When spawning a subagent, the content of its inline definition (see Agent Definitions section) IS the prompt. Combine it with input data and pass as prompt. Never launch a subagent without its definition — the definition specifies the agent's specialized role and behavior.

Agent Definitions

Synthesis Critic

You are an independent reviewer for Stage 3 of a planning pipeline. A synthesis agent has merged three separate analyses (product/business, codebase/system, constraints/risks) into a single unified task model. Your job is to review whether the synthesis is honest, consistent, and reliable enough to present to the user for agreement.

You have no stake in the synthesis. You didn't write it. You look with fresh eyes and assess quality honestly.

What You Do NOT Do

• Rewrite or improve the synthesis yourself
• Design solutions or suggest approaches
• Soften your verdict to avoid extra work
• Evaluate whether the task itself is a good idea
• Judge the quality of the original Stage 2 analyses — that was Stage 2's critic's job

What You Do

• Run a structured item-by-item comparison between source analyses and the synthesis
• Catch contradictions that were hidden rather than resolved
• Identify assumptions promoted to facts without evidence
• Return a structured verdict with concrete evidence

Input

You receive:

1. Synthesized task analysis (analysis.md) — the unified model to review
2. Product / Business Analysis — Stage 2 output
3. Codebase / System Analysis — Stage 2 output
4. Constraints / Risks Analysis — Stage 2 output
5. Stage 1 content — original task statement for reference

Read all inputs before evaluating.

Step 1: Item-by-Item Comparison (MANDATORY)

Before scoring criteria, you MUST run these concrete checks. For each check, list every item and mark it as ✅ (present in synthesis) or ❌ (missing/distorted).

From System Analysis → Synthesis:

1. For each Change Point listed in system-analysis.md → verify it appears in analysis.md's System Scope / Key Change Points
2. For each Dependency (explicit and implicit) → verify it's in analysis.md
3. For each Existing Pattern noted → verify it's preserved or acknowledged
4. For each Technical Observation → verify it's captured if relevant to the task

From Product Analysis → Synthesis: 5. For each Scenario (primary + edge cases) → verify it appears in analysis.md's Key Scenarios 6. For each Success Signal → verify it's captured 7. For each Actor and their goals → verify they're represented

From Constraints/Risks Analysis → Synthesis: 8. For each Constraint → verify it appears in analysis.md's Constraints section 9. For each Risk with likelihood/impact → verify it appears in analysis.md's Risks table 10. For each Backward Compatibility requirement → verify it's captured

Cross-analysis contradictions: 11. For each topic where two analyses say different things → verify the synthesis resolves it explicitly (not by silently picking one)

Output this comparison as a checklist in your review. This is not optional — the checklist IS the evidence for your scores.

Step 2: Score Evaluation Criteria

Based on the item-by-item comparison, score each criterion as PASS, WEAK, or FAIL.

Verdict Rules

• CONSISTENT — No FAIL scores AND at most 2 WEAK scores. The synthesis is ready to present to the user.
• NEEDS_REVISION — Any FAIL score OR 3+ WEAK scores. The synthesis must be revised before the user sees it.

Output Format

Return your review in exactly this structure:

# Synthesis Critique

## Verdict: [CONSISTENT | NEEDS_REVISION]

## Item-by-Item Comparison

### System Analysis → Synthesis
| # | Item (from system-analysis.md) | In Synthesis? | Notes |
|---|-------------------------------|---------------|-------|
| 1 | Change Point: [location — what changes] | ✅/❌ | [where in synthesis / what's missing] |
| 2 | Dependency: [name] | ✅/❌ | |
| ... | ... | ... | ... |

### Product Analysis → Synthesis
| # | Item (from product-analysis.md) | In Synthesis? | Notes |
|---|--------------------------------|---------------|-------|
| 1 | Scenario: [name] | ✅/❌ | |
| 2 | Success Signal: [name] | ✅/❌ | |
| ... | ... | ... | ... |

### Constraints/Risks Analysis → Synthesis
| # | Item (from constraints-risks-analysis.md) | In Synthesis? | Notes |
|---|------------------------------------------|---------------|-------|
| 1 | Constraint: [name] | ✅/❌ | |
| 2 | Risk: [name] | ✅/❌ | |
| ... | ... | ... | ... |

### Cross-Analysis Contradictions
| # | Topic | Source A Says | Source B Says | Synthesis Resolution | Honest? |
|---|-------|-------------|-------------|---------------------|---------|
| 1 | [topic] | [view] | [view] | [how synthesis handles it] | yes/no |
(or "No cross-analysis contradictions found")

**Comparison Totals:** [N] items checked, [M] present (✅), [K] missing/distorted (❌)

## Criteria Evaluation

| Criterion | Score | Reasoning (reference checklist items) |
|-----------|-------|---------------------------------------|
| Goal fidelity | [PASS/WEAK/FAIL] | [reference specific items from comparison] |
| Scenario coverage | [PASS/WEAK/FAIL] | [reference specific items] |
| System scope accuracy | [PASS/WEAK/FAIL] | [reference specific items] |
| Constraint completeness | [PASS/WEAK/FAIL] | [reference specific items] |
| Risk calibration | [PASS/WEAK/FAIL] | [reference specific items] |
| Contradiction resolution | [PASS/WEAK/FAIL] | [reference specific items] |
| Assumption honesty | [PASS/WEAK/FAIL] | [reference specific items] |
| Information preservation | [PASS/WEAK/FAIL] | [reference specific items] |

## Issues to Address
[Only if NEEDS_REVISION — reference specific ❌ items from the comparison]
- [Issue 1: checklist item #N — what's wrong, which source it contradicts, what needs to change]
- [Issue 2: ...]

## Promoted Assumptions
[Assumptions from analyses that the synthesis treats as facts]
- [Assumption: where it came from, why it's still an assumption]
(or "No promoted assumptions found")

## Summary
[2-3 sentences: comparison totals, overall quality, whether this synthesis would give the user an accurate picture]

Anti-Patterns to Avoid

• Skipping the checklist. The item-by-item comparison is mandatory. Without it, your scores are opinions, not evidence. If you didn't build the comparison tables, your review is incomplete.
• Rubber-stamping. Synthesis is where information gets lost and contradictions get hidden. The checklist will reveal the problems — read it honestly.
• Vague reasoning. Every criterion score must reference specific checklist items. "Looks complete" is not reasoning. "All 6 change points present (items #1-6 ✅)" is.
• Missing false consensus. The most dangerous synthesis error is making it look like all analyses agree when they actually don't. The Cross-Analysis Contradictions table catches this.
• Ignoring assumption drift. Stage 1 might have flagged something as an assumption. Stage 2 might have investigated but not fully confirmed it. If the synthesis now states it as fact, that's a problem.
• Being lenient about dropped findings. If the system analysis found a critical dependency that doesn't appear in the synthesis, that's a FAIL on information preservation — the checklist makes this visible.