Agent Prompts & AGENTS.md

Saga can generate prompts tailored for coding agents (Claude Code, Copilot, Cursor, etc.) directly from your stories.

Per-story agent prompts

Right-click a story → Generate Agent Prompt. Saga:

1. Detects your stack (reads package.json, pyproject.toml, Cargo.toml, go.mod)
2. Scores workspace files for relevance to the story (filename/path overlap with the story's title and labels)
3. Shows a checklist of the most relevant files (with an estimated token count) so you can pick which to include as full-content context
4. Generates a Markdown prompt combining the story, stack info, directory layout, and selected file contents

The prompt opens in an editable panel — one tab per story — with Copy to Clipboard and Save actions. Nothing is written to disk until you explicitly save, to .saga/prompts/STORY-NNN.prompt.md.

AGENTS.md

Run Generate AGENTS.md (Command Palette, or the toolbar icon in the Epics & Stories view) to generate a repo-wide AGENTS.md — a single file describing your stack, layout, and conventions for any coding agent operating in the repo.

• If AGENTS.md already exists, its content is passed to the LLM for continuity, and shown in a diff-style comparison
• A .saga/AGENTS.md.lock file stores a hash of the last generated content — if the file was hand-edited since, you're warned before overwriting
• Accept writes the file and updates the lock; Regenerate re-calls the LLM; Discard closes without writing

Relevance scoring (v1 scope)

File relevance uses simple heuristic string-overlap scoring — no embeddings. Files score 0–1; anything above 0.1 is included, capped at the top 20. node_modules, .git, dist, build, out, .saga, and .vscode are always excluded.