agent-bober 0.9.0 → 0.10.1

package/dist/state/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/state/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,SAAS,EAAE,MAAM,cAAc,CAAC;AAEzC,OAAO,EACL,YAAY,EACZ,YAAY,EACZ,aAAa,EACb,cAAc,GACf,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EACL,QAAQ,EACR,QAAQ,EACR,cAAc,EACd,SAAS,GACV,MAAM,iBAAiB,CAAC;AAEzB,OAAO;AACL,cAAc;AACd,WAAW,EACX,kBAAkB;AAIlB,YAAY;AACZ,aAAa,EACb,WAAW,EACX,cAAc,GACf,MAAM,cAAc,CAAC;AAEtB,OAAO,EACL,YAAY,EACZ,YAAY,EACZ,YAAY,GACb,MAAM,qBAAqB,CAAC;AAE7B,OAAO,EACL,UAAU,EACV,UAAU,GACX,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EACL,WAAW,EACX,WAAW,GACZ,MAAM,oBAAoB,CAAC;AAE5B,MAAM,SAAS,GAAG,QAAQ,CAAC;AAC3B,MAAM,OAAO,GAAG,CAAC,WAAW,EAAE,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,CAAU,CAAC;AAEnF;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,WAAmB;IACtD,MAAM,SAAS,GAAG,IAAI,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC;IAC/C,MAAM,SAAS,CAAC,SAAS,CAAC,CAAC;IAE3B,KAAK,MAAM,GAAG,IAAI,OAAO,EAAE,CAAC;QAC1B,MAAM,SAAS,CAAC,IAAI,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC,CAAC;IACxC,CAAC;AACH,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/state/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,SAAS,EAAE,MAAM,cAAc,CAAC;AAEzC,OAAO,EACL,YAAY,EACZ,YAAY,EACZ,aAAa,EACb,cAAc,GACf,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EACL,QAAQ,EACR,QAAQ,EACR,cAAc,EACd,SAAS,GACV,MAAM,iBAAiB,CAAC;AAEzB,OAAO;AACL,cAAc;AACd,WAAW,EACX,kBAAkB;AAIlB,YAAY;AACZ,aAAa,EACb,WAAW,EACX,cAAc,GACf,MAAM,cAAc,CAAC;AAEtB,OAAO,EACL,YAAY,EACZ,YAAY,EACZ,YAAY,GACb,MAAM,qBAAqB,CAAC;AAE7B,OAAO,EACL,UAAU,EACV,UAAU,GACX,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EACL,WAAW,EACX,WAAW,GACZ,MAAM,oBAAoB,CAAC;AAE5B,OAAO,EACL,gBAAgB,EAChB,gBAAgB,EAChB,OAAO,EACP,QAAQ,EACR,iBAAiB,GAClB,MAAM,sBAAsB,CAAC;AAE9B,MAAM,SAAS,GAAG,QAAQ,CAAC;AAC3B,MAAM,OAAO,GAAG,CAAC,WAAW,EAAE,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,EAAE,cAAc,CAAU,CAAC;AAEnG;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,WAAmB;IACtD,MAAM,SAAS,GAAG,IAAI,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC;IAC/C,MAAM,SAAS,CAAC,SAAS,CAAC,CAAC;IAE3B,KAAK,MAAM,GAAG,IAAI,OAAO,EAAE,CAAC;QAC1B,MAAM,SAAS,CAAC,IAAI,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC,CAAC;IACxC,CAAC;AACH,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-bober",
-  "version": "0.9.0",
+  "version": "0.10.1",
   "description": "Generator-Evaluator multi-agent harness for building applications autonomously with any LLM. Supports Claude, GPT, Gemini, Ollama. Includes MCP server for Cursor/Windsurf.",
   "type": "module",
   "main": "dist/index.js",

package/skills/bober.architect/SKILL.md

@@ -0,0 +1,418 @@
+---
+name: bober.architect
+description: Run the 5-checkpoint architecture discussion flow — problem framing, approach selection, component design, integration strategy, and final assembly. Produces an architecture document with ADRs saved to .bober/architecture/.
+argument-hint: <feature-or-spec-description>
+handoffs:
+  - label: "Plan Feature"
+    command: /bober-plan
+    prompt: "Plan the architecture as sprint contracts"
+  - label: "Full Autonomous Run"
+    command: /bober-run
+    prompt: "Run all sprints from the plan autonomously"
+---
+
+# bober.architect — Architecture Skill
+
+You are running the **bober.architect** skill. Your job is to orchestrate a structured 5-checkpoint architecture discussion, producing an architecture document and Architecture Decision Records (ADRs). You do NOT write application code.
+
+You use the **Agent tool** to spawn subagents at each checkpoint. Between checkpoints, you present output to the user and wait for review before proceeding.
+
+## Step 1: Read Project Configuration
+
+Read `bober.config.json` from the project root. If it does not exist, tell the user to run `/bober-plan` first to initialize the project.
+
+Read `.bober/principles.md` if it exists — include relevant principles in subagent prompts.
+
+## Step 2: Generate Architecture ID
+
+Generate a unique architecture ID using the format: `arch-<YYYYMMDD>-<slug>`
+
+Where `<slug>` is derived from the first 5 words of the feature description, lowercased, spaces replaced with hyphens, special characters removed.
+
+Example: For "Payment processing service with retry logic", the ID is `arch-20260331-payment-processing-service-with-retry`.
+
+## Step 3: Check for Existing Research
+
+Check `.bober/research/` for any research document related to this spec. Look for files matching `*<keyword>*-research.md` where keywords come from the feature description. If found, read it and include the research findings in the Checkpoint 1 subagent prompt as additional context.
+
+## Step 4: Read Agent Definition
+
+Read `agents/bober-architect.md`. Include this as the system prompt for each checkpoint subagent.
+
+## Step 5: Checkpoint 1 — Problem Framing
+
+Spawn a subagent with this prompt structure:
+
+```
+You are the Bober Architect agent running Checkpoint 1: Problem Framing.
+
+## Architecture ID
+<ARCH_ID>
+
+## Feature Description
+<FEATURE_DESCRIPTION>
+
+## Research Context (if available)
+<RESEARCH_FINDINGS or "No prior research available.">
+
+## Project Root
+<PROJECT_ROOT>
+
+## Agent Definition
+<BOBER_ARCHITECT_AGENT_DEFINITION>
+
+## Your Task
+
+Run Checkpoint 1 (Problem Framing) from your agent definition.
+
+Ask and self-answer the 6 Checkpoint 1 questions by reading the codebase. For each question:
+1. Use Glob/Grep/Read to find evidence in the codebase
+2. Answer the question directly based on evidence
+3. If no codebase evidence exists for a constraint, state the assumption explicitly
+
+Produce the Problem Statement section exactly as specified in your agent definition.
+
+## Output Format
+
+Respond with:
+{
+  "checkpoint": 1,
+  "problemStatement": "<full markdown Problem Statement section>",
+  "assumptions": ["<assumption 1>", "..."],
+  "questionsAnswered": <number>
+}
+```
+
+Wait for the subagent to complete. Parse its output.
+
+Present the Problem Statement to the user:
+
+```
+## Checkpoint 1 Complete: Problem Framing
+
+<problemStatement content>
+
+---
+
+**Review this checkpoint.** You can:
+- **(A) Approve and continue** to Checkpoint 2: Approach Selection
+- **(B) Challenge specific constraints** — tell me what is wrong or missing
+- **(C) Restart this checkpoint** with different framing
+
+What would you like to do?
+```
+
+Wait for user response. If (B) or (C): respawn the subagent with the user's feedback included in the prompt. Repeat until the user approves.
+
+## Step 6: Checkpoint 2 — Approach Selection
+
+Spawn a subagent with the approved Problem Statement included:
+
+```
+You are the Bober Architect agent running Checkpoint 2: Approach Selection.
+
+## Architecture ID
+<ARCH_ID>
+
+## Problem Statement (approved at Checkpoint 1)
+<APPROVED_PROBLEM_STATEMENT>
+
+## Project Root
+<PROJECT_ROOT>
+
+## Agent Definition
+<BOBER_ARCHITECT_AGENT_DEFINITION>
+
+## Your Task
+
+Run Checkpoint 2 (Approach Selection) from your agent definition.
+
+CRITICAL RULES:
+- Present exactly 2 or 3 approaches. Never 1, never 4+.
+- Use the structured format from your agent definition — NOT paragraphs.
+- Each approach must have: Description (1 sentence), Pros (bullets), Cons (bullets), Best-for (1 sentence).
+- Select one approach with rationale citing Checkpoint 1 constraints.
+- Draft an ADR for this decision (do not save yet — the orchestrator will save it).
+
+## Output Format
+
+Respond with:
+{
+  "checkpoint": 2,
+  "approachesSection": "<full markdown Approaches section>",
+  "selectedApproach": "<Approach A/B/C name>",
+  "selectionRationale": "<1 sentence rationale>",
+  "adrDraft": "<full ADR-1 markdown content>"
+}
+```
+
+Present the approaches to the user:
+
+```
+## Checkpoint 2 Complete: Approach Selection
+
+<approachesSection content>
+
+---
+
+**Review this checkpoint.** You can:
+- **(A) Approve the selected approach** and continue to Checkpoint 3: Component Design
+- **(B) Select a different approach** — tell me which one and why
+- **(C) Request a different set of approaches** — describe what you want to see instead
+
+What would you like to do?
+```
+
+Wait for user response. Handle (B)/(C) by respawning with feedback.
+
+## Step 7: Checkpoint 3 — Component Design
+
+Spawn a subagent with Problem Statement and selected approach:
+
+```
+You are the Bober Architect agent running Checkpoint 3: Component Design.
+
+## Architecture ID
+<ARCH_ID>
+
+## Problem Statement (Checkpoint 1)
+<APPROVED_PROBLEM_STATEMENT>
+
+## Selected Approach (Checkpoint 2)
+<SELECTED_APPROACH_NAME>: <SELECTION_RATIONALE>
+
+## Project Root
+<PROJECT_ROOT>
+
+## Agent Definition
+<BOBER_ARCHITECT_AGENT_DEFINITION>
+
+## Your Task
+
+Run Checkpoint 3 (Component Design) from your agent definition.
+
+CRITICAL RULES:
+- TypeScript-style interfaces are mandatory for every component. No prose descriptions.
+- Responsibility is exactly 1 sentence. No more.
+- List component dependencies explicitly.
+- Read the actual codebase to ground interface designs in existing patterns.
+- Draft ADRs for any significant component boundary decisions.
+
+## Output Format
+
+Respond with:
+{
+  "checkpoint": 3,
+  "componentBreakdownSection": "<full markdown Component Breakdown section>",
+  "componentCount": <number>,
+  "adrDrafts": ["<ADR-2 markdown>", "<ADR-3 markdown if needed>"]
+}
+```
+
+Present components to user:
+
+```
+## Checkpoint 3 Complete: Component Design
+
+<componentBreakdownSection content>
+
+---
+
+**Review this checkpoint.** You can:
+- **(A) Approve and continue** to Checkpoint 4: Integration Strategy
+- **(B) Challenge component boundaries** — tell me which components need redesigning
+- **(C) Request different interface signatures** — specify what should change
+
+What would you like to do?
+```
+
+Handle (B)/(C) by respawning with feedback.
+
+## Step 8: Checkpoint 4 — Integration Strategy
+
+Spawn a subagent with all prior outputs:
+
+```
+You are the Bober Architect agent running Checkpoint 4: Integration Strategy.
+
+## Architecture ID
+<ARCH_ID>
+
+## Problem Statement (Checkpoint 1)
+<APPROVED_PROBLEM_STATEMENT>
+
+## Selected Approach (Checkpoint 2)
+<SELECTED_APPROACH>
+
+## Component Breakdown (Checkpoint 3)
+<APPROVED_COMPONENT_BREAKDOWN>
+
+## Project Root
+<PROJECT_ROOT>
+
+## Agent Definition
+<BOBER_ARCHITECT_AGENT_DEFINITION>
+
+## Your Task
+
+Run Checkpoint 4 (Integration Strategy) from your agent definition.
+
+CRITICAL RULES:
+- Data flow must be a concrete call chain (A → B.method(x) → C.method(y)), not an abstract description.
+- Every integration risk must have: Risk, Severity (critical/high/medium/low), Mitigation.
+- API Contracts table must include Error Cases column.
+- Draft ADRs for key integration decisions.
+
+## Output Format
+
+Respond with:
+{
+  "checkpoint": 4,
+  "integrationStrategySection": "<full markdown Integration Strategy section>",
+  "riskCount": <number of risks identified>,
+  "adrDrafts": ["<ADR markdown if needed>"]
+}
+```
+
+Present integration strategy to user:
+
+```
+## Checkpoint 4 Complete: Integration Strategy
+
+<integrationStrategySection content>
+
+---
+
+**Review this checkpoint.** You can:
+- **(A) Approve and continue** to Checkpoint 5: Final Assembly
+- **(B) Challenge the data flow** — describe what is wrong or missing
+- **(C) Add integration risks** — tell me what risks were missed
+
+What would you like to do?
+```
+
+Handle (B)/(C) by respawning with feedback.
+
+## Step 9: Checkpoint 5 — Final Assembly
+
+Spawn a subagent to compile the complete architecture document:
+
+```
+You are the Bober Architect agent running Checkpoint 5: Final Assembly.
+
+## Architecture ID
+<ARCH_ID>
+
+## Feature Description
+<FEATURE_DESCRIPTION>
+
+## All Checkpoint Outputs
+
+### Problem Statement (Checkpoint 1)
+<APPROVED_PROBLEM_STATEMENT>
+
+### Approaches (Checkpoint 2)
+<APPROVED_APPROACHES_SECTION>
+
+### Component Breakdown (Checkpoint 3)
+<APPROVED_COMPONENT_BREAKDOWN>
+
+### Integration Strategy (Checkpoint 4)
+<APPROVED_INTEGRATION_STRATEGY>
+
+## ADR Drafts
+<ALL_ADR_DRAFTS from prior checkpoints>
+
+## Project Root
+<PROJECT_ROOT>
+
+## Agent Definition
+<BOBER_ARCHITECT_AGENT_DEFINITION>
+
+## Your Task
+
+Compile the complete architecture document per the output format in your agent definition.
+
+CRITICAL RULES:
+- The complete document MUST NOT exceed 500 lines.
+- Each ADR MUST NOT exceed 50 lines.
+- Executive Summary is exactly 3-5 sentences — no filler.
+- Number ADRs sequentially: ADR-1, ADR-2, etc.
+- Include ALL 10 required sections in the exact order specified.
+- If content exceeds 500 lines, distill — remove filler, shorten descriptions, combine redundant sections.
+- Save the architecture document to .bober/architecture/<id>-architecture.md
+- Save each ADR to .bober/architecture/<id>-adr-<N>.md (separate files)
+- Ensure .bober/architecture/ directory exists before writing.
+
+## Required Document Sections (in this order)
+1. Executive Summary
+2. Problem Statement
+3. System Overview
+4. Component Breakdown
+5. Data Model
+6. API Contracts
+7. Integration Strategy
+8. Architecture Decision Records
+9. Risk Assessment
+10. Open Questions
+
+## Output Format
+
+After saving all files, respond with:
+{
+  "checkpoint": 5,
+  "architectureId": "<ARCH_ID>",
+  "documentPath": ".bober/architecture/<id>-architecture.md",
+  "adrPaths": [".bober/architecture/<id>-adr-1.md", "..."],
+  "componentCount": <number>,
+  "decisionCount": <number of ADRs>,
+  "documentLineCount": <actual line count>,
+  "summary": "<2-3 sentence summary of the architecture>"
+}
+```
+
+Wait for the subagent to complete.
+
+## Step 10: Save History and Present Final Output
+
+Append to `.bober/history.jsonl`:
+
+```json
+{"event":"architecture-completed","architectureId":"...","feature":"...","componentCount":N,"decisionCount":N,"timestamp":"..."}
+```
+
+Present the final output to the user:
+
+```
+## Architecture Complete: <feature description>
+
+**Architecture ID:** <arch_id>
+**Document:** .bober/architecture/<id>-architecture.md
+**ADRs:** <list each ADR path>
+
+### Summary
+<summary from Checkpoint 5>
+
+### Components (<count>)
+<list component names>
+
+### Decisions (<count> ADRs)
+<list ADR titles>
+
+### Next Steps
+- Run `/bober-plan <feature-description>` to decompose this architecture into sprint contracts
+- Run `/bober-sprint` to start building
+```
+
+## Error Handling
+
+- If a subagent produces invalid JSON: retry once with the same prompt, then present the raw output to the user and ask them to proceed manually
+- If the architecture directory cannot be created: report the filesystem error clearly
+- If the user rejects a checkpoint 3+ times: ask if they want to restart from Checkpoint 1 with different framing
+- If the feature description argument is empty: tell the user the skill requires a feature description argument
+
+## Important: Checkpoint Isolation
+
+Each checkpoint subagent receives only the outputs from PRIOR approved checkpoints — not the raw outputs of ongoing checkpoints. This prevents circular reasoning and ensures each checkpoint builds on approved decisions.
+
+The skill (this document) is the coordinator. Subagents produce; the skill assembles.

package/skills/bober.research/SKILL.md

@@ -0,0 +1,219 @@
+---
+name: bober.research
+description: Run the two-phase research process for a feature — generates exploration questions then explores the codebase to produce a factual research document.
+argument-hint: <feature-description>
+handoffs:
+  - label: "Plan Feature"
+    command: /bober-plan
+    prompt: "Plan the researched feature"
+---
+
+# bober.research — Research Skill
+
+You are running the **bober.research** skill. Your job is to orchestrate the two-phase research process for a feature description, producing a factual codebase research document that can inform planning.
+
+You do NOT implement code. You do NOT plan sprints. You spawn subagents using the **Agent tool** and coordinate their outputs.
+
+## Step 1: Read Project Configuration
+
+Read `bober.config.json` from the project root. If it does not exist, tell the user to run `/bober-plan` first to initialize the project.
+
+Read `.bober/principles.md` if it exists. Include it in subagent prompts.
+
+## Step 2: Generate a Research ID
+
+Generate a unique research ID using the format: `research-<YYYYMMDD>-<slug>`
+
+Where `<slug>` is derived from the first 5 words of the feature description, lowercased, spaces replaced with hyphens, special characters removed.
+
+Example: For "Add user authentication with email", the slug is `add-user-authentication-with-email`, giving ID `research-20260331-add-user-authentication-with-email`.
+
+## Step 3: Phase 1 — Question Generation
+
+Spawn a subagent using the Agent tool with this prompt structure:
+
+```
+You are the Bober Researcher agent, Phase 1: Question Generation.
+
+## Your Task
+
+Given the feature description below, generate 5–8 specific exploration questions that will guide codebase exploration. These questions will be passed to a SEPARATE agent that has NO knowledge of the feature — so your questions must be self-contained and answerable by reading the codebase.
+
+## Rules
+
+- Generate ONLY questions, no preamble, no explanation
+- Questions must be specific to what a developer would need to explore
+- Questions must be answerable by reading files (not by building or running code)
+- Do NOT suggest implementations or make recommendations
+- Do NOT explore the codebase yourself — just generate questions
+
+## Feature Description
+
+<FEATURE_DESCRIPTION>
+
+## Project Root
+
+<PROJECT_ROOT>
+
+## Agent Definition
+
+<BOBER_RESEARCHER_SYSTEM_PROMPT>
+
+## Output Format
+
+Respond with ONLY a JSON array of question strings. No markdown fences, no explanation.
+```
+
+Wait for the Phase 1 subagent to complete. Parse its output as a JSON array of strings.
+
+**If Phase 1 fails or produces fewer than 3 questions:** Log the error and abort. Tell the user what went wrong.
+
+## Step 4: Phase 2 — Codebase Exploration
+
+Spawn a second subagent using the Agent tool. This subagent receives ONLY the questions, NOT the feature description. This isolation is mandatory.
+
+Use this prompt structure:
+
+```
+You are the Bober Researcher agent, Phase 2: Codebase Exploration.
+
+## Your Task
+
+You have been given a list of exploration questions. Your job is to explore the codebase and answer each question with factual findings. You do NOT know what feature is being built — this is intentional to prevent bias in your research.
+
+## Exploration Questions
+
+<QUESTIONS_FROM_PHASE_1>
+
+## Research ID
+
+<RESEARCH_ID>
+
+## Project Root
+
+<PROJECT_ROOT>
+
+## Agent Definition
+
+<BOBER_RESEARCHER_SYSTEM_PROMPT>
+
+## Instructions
+
+1. Work through each question systematically using Read, Grep, and Glob tools
+2. Record exact file paths and line numbers for every finding
+3. Produce a factual research document with ONLY these sections:
+   - Architecture Overview
+   - Existing Patterns (with file:line refs)
+   - Key Files
+   - Integration Points
+   - Test Coverage
+   - Risk Areas
+4. No recommendations. No opinions. Facts only.
+
+## Output Format
+
+Respond with a JSON object:
+{
+  "researchId": "<research ID>",
+  "sections": {
+    "architectureOverview": "<string>",
+    "existingPatterns": "<string>",
+    "keyFiles": "<string>",
+    "integrationPoints": "<string>",
+    "testCoverage": "<string>",
+    "riskAreas": "<string>"
+  },
+  "filesExplored": ["<list of file paths you read>"],
+  "questionsAnswered": <number>
+}
+
+No markdown fences, no explanation. Only the JSON object.
+```
+
+Wait for the Phase 2 subagent to complete.
+
+## Step 5: Save Research Document
+
+Save the research output to `.bober/research/<researchId>-research.md` using this markdown format:
+
+```markdown
+# Research: <feature description (first 80 chars)>
+
+**Research ID:** <researchId>
+**Generated:** <ISO-8601 timestamp>
+**Questions Explored:** <count>
+**Files Explored:** <count>
+
+---
+
+## Architecture Overview
+
+<architectureOverview>
+
+## Existing Patterns
+
+<existingPatterns>
+
+## Key Files
+
+<keyFiles>
+
+## Integration Points
+
+<integrationPoints>
+
+## Test Coverage
+
+<testCoverage>
+
+## Risk Areas
+
+<riskAreas>
+
+---
+
+*Generated by bober.research — factual findings only, no implementation recommendations.*
+```
+
+Ensure the `.bober/research/` directory exists before writing. Create it if needed.
+
+Also append to `.bober/history.jsonl`:
+```json
+{"event":"research-completed","researchId":"...","feature":"...","questionsGenerated":N,"filesExplored":N,"timestamp":"..."}
+```
+
+## Step 6: Output Summary
+
+Present a summary to the user:
+
+```
+## Research Complete: <feature description>
+
+**Research ID:** <researchId>
+**Saved to:** .bober/research/<researchId>-research.md
+
+### Questions Explored (<count>)
+1. <question 1>
+2. <question 2>
+...
+
+### Files Explored (<count>)
+<list key files>
+
+### Key Findings
+<3-5 bullet points with the most important factual findings>
+
+### Next Steps
+Run `/bober-plan <feature-description>` to use this research to create a plan.
+```
+
+## Error Handling
+
+- If Phase 1 produces invalid JSON: retry once with the same prompt, then abort
+- If Phase 2 produces invalid JSON: save whatever text was produced as a raw research note, log a warning, but do NOT fail silently
+- If the research directory cannot be created: report the filesystem error clearly
+- If the feature description argument is empty: tell the user the skill requires a feature description argument
+
+## Important: Two-Phase Isolation
+
+The feature description MUST NOT be passed to the Phase 2 subagent. This is the entire point of the two-phase process. Phase 2's research is unbiased because the exploring agent does not know what is being built — it only knows what to look for.