4.6 Reflection / Self-Refine

Generate → inspect → identify problems → revise → recheck. Create a monthly business report from real CSV data or refine an existing draft. Accept a result only after both model review and deterministic checks pass.

mermaid

flowchart TD
  L[Main Loop] --> S[Load request, CSV and Memory]
  S --> G[Generate draft / load user draft]
  G --> D[Check figures and source quotes]
  D --> R[Model review]
  R -->|issues| W[Revise complete draft]
  W --> D
  R -->|both gates pass| P[Publish verified analysis]
  R -->|limits / no progress / human| H[Preserve draft and review history]

The main Loop composes stage Graphs using graphStep; Graphs contain only public Worker nodes. Arrows show Loop decisions. Context uses Redis and Memory uses persistent SQLite. CSV, calculation, citation and artifact tools live outside Core.

Run

Use Node 24, real Redis, ditto.yaml and model credentials in the environment. Artifacts are saved under the ignored .examples-reflection-tasks/ directory.

export DITTO_WORKER_CONTEXT_REDIS_URL=redis://127.0.0.1:6379
npm run example:reflection -- --provider deepseek
npm run example:reflection -- --provider deepseek --scenario flawed-draft
npm run example:reflection -- --provider deepseek --scenario missing-data
npm run example:reflection -- --provider deepseek --stop-after review
npm run example:reflection -- --provider deepseek --directory .examples-reflection-tasks/cli-XXXXXX
npm run check:examples:reflection:package -- --provider deepseek

--rounds limits draft versions, --calls limits model attempts and --goal supplies the task goal. Checkpoints support draft|review|report; resume preserves the request. A generated draft may pass immediately; the seeded-draft scenario demonstrates revision of wrong calculations and missing citations.

API, runnable consumer and failure/recovery contracts · Application tools.

4.6 Reflection / Self-Refine ​

Run ​

4.6 Reflection / Self-Refine

Run