Worldshaper/docs/kb/queue-workflow.md

132 lines
3.6 KiB
Markdown

# Request Queue Workflow
This document describes a practical first pass for automated request analysis using the KB scaffold.
## Goal
Turn raw launcher submissions into structured request items that are:
- titled
- categorized
- tagged
- explained in plain language
- paired with a realistic implementation approach
## Recommended Pipeline
1. Read one pending submission.
2. Split it into candidate atomic requests.
3. Retrieve likely systems from `docs/kb/systems.json`.
4. Load the matching `docs/kb/systems/*.md` files.
5. Load the standardized request tags from `docs/kb/tags.json`.
6. Ask the local model for JSON that matches `docs/kb/request-analysis-schema.json`.
7. Validate the JSON.
8. Promote high-confidence items and hold low-confidence items for review.
## Suggested Confidence Rules
- `>= 0.85`
- auto-promote to active
- `0.65 - 0.84`
- store as needs review
- `< 0.65`
- keep as pending or send to manual triage
## Retrieval Hints
Search `systems.json` using:
- request text
- obvious UI words like "layer", "chunk", "overview", "graphics", "animation"
- aliases
- past tags from similar requests
If a submission touches multiple subsystems, feed the model the top two to four matching docs instead of the whole KB.
## Prompt Skeleton
Use a strict instruction similar to this:
```text
You are processing Worldshaper editor requests.
You must:
- split the submission into one or more atomic requests
- assign one primary category per item
- assign tags grounded in the provided KB and limited to `docs/kb/tags.json`
- write a short descriptive title
- explain the user intent in plain language
- propose a likely implementation path
- provide a short review rationale and possible options when confidence is low
- avoid inventing systems that are not in the KB
- return only valid JSON matching the provided schema
If you are unsure, lower confidence and use statusRecommendation = "needs_review".
```
Then provide:
- the raw submission
- the JSON schema
- the retrieved KB docs
## Good First Storage Shape
Keep the existing launcher request record, but add optional analysis metadata:
- `analysisStatus`
- `analysisConfidence`
- `analysisModel`
- `analysisCreatedAt`
- `analysisUpdatedAt`
- `analysisPayload`
This lets you keep the current user-facing `pending` and `active` states while adding a back-office workflow.
## First Implementation Target
The simplest useful worker would:
1. read pending requests from the launcher request API/store
2. analyze one request at a time
3. write structured analysis back to the same store
4. never delete the original `sourceText`
The current first-pass implementation is:
```bash
npm run analyze:requests
```
Useful environment variables:
- `REQUEST_ANALYZER_API_BASE`
- `REQUEST_ANALYZER_PROVIDER`
- `REQUEST_ANALYZER_MODEL_BASE_URL`
- `REQUEST_ANALYZER_MODEL`
- `REQUEST_ANALYZER_API_KEY`
- `REQUEST_ANALYZER_MAX_TOKENS`
- `REQUEST_ANALYZER_THINKING`
- `REQUEST_ANALYZER_PROMOTE_THRESHOLD`
- `DEEPSEEK_API_KEY`
DeepSeek path:
- set `DEEPSEEK_API_KEY`
- leave the provider on its default auto-detect setting
- the worker will use DeepSeek-hosted chat completions automatically
Behavior:
- high-confidence all-active results replace the pending submission with active request rows
- mixed or lower-confidence results stay on the pending submission as review metadata
- failures are stored as analysis errors instead of silently disappearing
## What Not To Automate First
- automatic deletion
- automatic deduplication with destructive merges
- direct code generation
- automatic schema migrations
The first win is dependable triage, not full autonomy.