conversation-lab
Self-improving conversation convergence loop for SPARTA stress tests. Diagnoses unsatisfied sessions, re-runs until personas say satisfactory, outputs structured JSON for /assess consumption. Use when conversation quality is low, sessions are partial, or you need to converge conversations.
Informations de source
- Dépôt
- grahama1970/agent-stack-public
- Dernière activité de la source
- 24 septembre 2026 à 15:51
- Langue détectée de SKILL.md
- anglais
- Étoiles
- 0
- Forks
- 0
Options d'installation
Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.
Vérifiez les fichiers source
Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.
Explorateur de fichiers
13 fichiersAffichage de SKILL.md
SKILL.md
Instructions source · Aperçu en lecture seule- name
- conversation-lab
- description
- Self-improving conversation convergence loop for SPARTA stress tests. Diagnoses unsatisfied sessions, re-runs until personas say satisfactory, outputs structured JSON for /assess consumption. Use when conversation quality is low, sessions are partial, or you need to converge conversations.
- triggers
- ["conversation lab","converge conversations","fix unsatisfied sessions","conversation convergence","diagnose conversations","rerun sessions"]
- metadata
- {"short-description":"Conversation convergence engine","version":"1.0.0"}
- provides
- ["conversation-convergence","session-diagnostics","turn-optimization"]
- composes
- ["sparta-stress-test","review-conversation","episodic-archiver","task-monitor","memory","agentic-evals"]
- taxonomy
- ["validation","iteration","convergence"]
- disciplines
- ["persona-simulation","evaluation-quality"]
> STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY ENDPOINT.
# Conversation Lab
Self-improving convergence loop for SPARTA stress test conversations.
Follows the Shadow-Lego pattern: run -> assess -> diagnose -> fix -> re-run.
## Commands
| Command | Description |
|---------|-------------|
| `diagnose [file]` | Structured JSON diagnosis for `/assess` (use `--format md` for Rich tables) |
| `report [file]` | Full markdown report with conversation transcripts (like `/tmp/conversation_flows_*.md`) |
| `data [file]` | Export JSON chart data for `/create-figure` (radar, heatmap, grade distribution) |
| `converge [file]` | Re-run unsatisfied sessions until personas are happy or ceiling hit |
| `optimize` | Analyze session data + episodic archives to recommend optimal turn counts |
| `status` | Show convergence state (running/complete/stalled) |
## Usage
```bash
# Diagnose existing sessions (JSON output)
./run.sh diagnose /path/to/sessions_*.jsonl
# Run convergence on unsatisfied sessions
./run.sh converge /path/to/sessions_*.jsonl --max-cycles 3
# Dry run (no LLM calls)
./run.sh converge /path/to/sessions_*.jsonl --dry-run
# Check turn optimization from episodic data
./run.sh optimize
# Human-readable markdown report (full transcripts)
./run.sh report /path/to/sessions_*.jsonl -o /tmp/conversation_flows.md
# Export chart data for /create-figure
./run.sh data /path/to/sessions_*.jsonl -o /tmp/charts/
# Then: cd .pi/skills/create-figure && ./run.sh radar --input /tmp/charts/radar.json
```
## Task-Monitor Integration
```bash
# Progress is reported automatically during converge
cat conversation_lab_task_state.json | jq
```
## NDJSON Streaming
```bash
./run.sh converge sessions.jsonl --json-stream | tee converge_results.jsonl
```
## Convergence Logic
1. **DIAGNOSE** — Read sessions, classify failures (zero QRA, over-clarification, regression, wrong answer, coverage gap)
2. **FILTER** — Select rerun candidates (unsatisfied + rerun_eligible)
3. **RE-RUN** — Call sparta-stress-test with same seeds, higher CONVO_MAX_ROUNDS
4. **COMPARE** — Grade delta between original and new session
5. **ARCHIVE** — Feed improvements to /episodic-archiver
6. **DECIDE** — Converging? Continue. Plateau? Stop. Regressing? Rollback.
### Stopping Conditions
- `satisfied_rate >= 0.80` — 80% persona satisfaction
- `max_cycles` exhausted (default 5)
- Plateau: 2 consecutive cycles with < 5% improvement
- Budget: total LLM calls exceed threshold
Voir sur GitHub