29f139e1cd
RFC 0012 was implemented with Option B (MCP orchestrates via Ollama) instead of Option A (Claude orchestrates via Task tool). This caused: - No parallel agents spawned - Fake Ollama responses instead of real deliberation - Inline JSON instead of .dialogue.md files Fix by removing blue_alignment_play tool entirely. Claude now orchestrates alignment dialogues directly using Task tool per ADR 0014. Also: - Add pub mod resources for RFC 0016/0017 (was missing) - Update lib.rs threading for spawn_blocking - Add .blue/worktrees/ to gitignore - Update database schema Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
141 lines
5.7 KiB
Markdown
141 lines
5.7 KiB
Markdown
# Spike: Alignment Dialogue Architecture Mismatch
|
|
|
|
**Date:** 2025-01-24
|
|
**Time-box:** 45 minutes
|
|
**Status:** Complete
|
|
|
|
## Problem
|
|
|
|
User invoked alignment dialogue expecting:
|
|
- Background agents spawned in parallel
|
|
- `.dialogue` format output
|
|
- Multi-round convergence tracking
|
|
|
|
Got instead:
|
|
- Inline text response (no agents spawned)
|
|
- No file created
|
|
- No actual expert deliberation
|
|
|
|
## Root Cause
|
|
|
|
**Blue's RFC 0012 implementation is architecturally wrong.**
|
|
|
|
Blue created `blue_alignment_play` as an MCP tool that:
|
|
- Runs synchronously in the MCP server
|
|
- Uses Ollama to generate fake "expert" responses
|
|
- Outputs JSON directly
|
|
|
|
**Coherence-MCP worked differently:**
|
|
- MCP server provides **helper tools only** (extract, lint, validate)
|
|
- **Claude orchestrates the dialogue itself** using the Task tool
|
|
- Spawns N parallel background agents in a single message
|
|
- Collects outputs from JSONL files
|
|
- Updates `.dialogue.md` file between rounds
|
|
|
|
## The Correct Architecture (from coherence-mcp ADR 0006)
|
|
|
|
```
|
|
┌─────────────────────────────────────────────────────────┐
|
|
│ CLAUDE SESSION (💙 Judge) │
|
|
│ │
|
|
│ 1. Recognize "play alignment" request │
|
|
│ 2. Spawn N agents IN PARALLEL (single message) │
|
|
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
|
|
│ │🧁 Agent1│ │🧁 Agent2│ │🧁 Agent3│ │🧁 Agent4│ │
|
|
│ └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘ │
|
|
│ │ │ │ │ │
|
|
│ ▼ ▼ ▼ ▼ │
|
|
│ ┌─────────────────────────────────────────────┐ │
|
|
│ │ /tmp/claude/{session}/tasks/{id}.output │ │
|
|
│ └─────────────────────────────────────────────┘ │
|
|
│ │ │
|
|
│ 3. Extract outputs via alignment_extract_dialogue │
|
|
│ 4. Score responses, update .dialogue.md │
|
|
│ 5. Repeat until convergence │
|
|
│ │
|
|
│ MCP TOOLS (helpers only): │
|
|
│ - alignment_extract_dialogue: read agent JSONL │
|
|
│ - alignment_dialogue_lint: validate format │
|
|
│ - alignment_dialogue_save: persist to docs │
|
|
└─────────────────────────────────────────────────────────┘
|
|
```
|
|
|
|
## Key Differences
|
|
|
|
| Aspect | Coherence-MCP (Correct) | Blue (Current) |
|
|
|--------|-------------------------|----------------|
|
|
| Orchestration | Claude + Task tool | MCP server |
|
|
| Agent spawning | Parallel background agents | None (fake inline) |
|
|
| LLM calls | Each agent is a Claude instance | Ollama in MCP |
|
|
| Output format | `.dialogue.md` file | JSON response |
|
|
| Multi-round | Real convergence loop | Single response |
|
|
| Judge role | Claude session | N/A |
|
|
|
|
## What Needs to Change
|
|
|
|
### 1. Delete `blue_alignment_play`
|
|
|
|
The tool that tries to run dialogues in MCP is wrong. Remove it.
|
|
|
|
### 2. Add Helper Tools Only
|
|
|
|
```rust
|
|
// Extract text from spawned agent JSONL
|
|
blue_extract_dialogue(task_id: str) -> String
|
|
|
|
// Validate dialogue format
|
|
blue_dialogue_lint(file_path: str) -> LintResult
|
|
|
|
// Save dialogue to .blue/docs/dialogues/
|
|
blue_dialogue_save(title: str, content: str) -> Result
|
|
```
|
|
|
|
### 3. Add ADR 0006 to Blue
|
|
|
|
Copy from coherence-mcp:
|
|
- `.blue/docs/adrs/0006-alignment-dialogue-agents.md`
|
|
|
|
This teaches Claude:
|
|
- N+1 agent architecture
|
|
- Spawning pattern (parallel Tasks)
|
|
- Convergence criteria
|
|
- File format
|
|
|
|
### 4. Reference in CLAUDE.md
|
|
|
|
Add to Blue's CLAUDE.md:
|
|
```markdown
|
|
## Alignment Dialogues
|
|
|
|
When asked to "play alignment" or run expert deliberation:
|
|
- See ADR 0006 for the N+1 agent architecture
|
|
- Use Task tool to spawn parallel agents
|
|
- Collect outputs via blue_extract_dialogue
|
|
- Update .dialogue.md file between rounds
|
|
```
|
|
|
|
### 5. Create Skill (Optional)
|
|
|
|
A `/alignment` skill could encapsulate the pattern, but the core behavior comes from Claude understanding ADR 0006.
|
|
|
|
## Files to Copy from Coherence-MCP
|
|
|
|
1. `docs/adrs/0006-alignment-dialogue-agents.md` → `.blue/docs/adrs/`
|
|
2. `docs/patterns/alignment-dialogue-pattern.md` → `.blue/docs/patterns/`
|
|
3. Handler logic from `crates/alignment-mcp/src/handlers/dialogue.rs`
|
|
4. Lint logic from `crates/alignment-mcp/src/handlers/dialogue_lint.rs`
|
|
|
|
## Immediate Action
|
|
|
|
The current `blue_alignment_play` tool should be removed. It gives the illusion of working but doesn't actually spawn agents or create proper dialogues.
|
|
|
|
The ADR that was already copied to Blue (`.blue/docs/adrs/0006-alignment-dialogue-agents.md`) needs to be referenced so Claude knows the pattern.
|
|
|
|
## Outcome
|
|
|
|
Blue's alignment dialogue feature was implemented wrong. The MCP server should provide extraction/validation tools, not orchestration. Claude itself orchestrates using the Task tool to spawn parallel agents. Fixing this requires:
|
|
|
|
1. Removing `blue_alignment_play`
|
|
2. Adding `blue_extract_dialogue` helper
|
|
3. Ensuring ADR 0006 is in Claude's context
|
|
4. Optionally creating a `/alignment` skill
|