feat: add Project entity with auto-injected agent context

[wesbillman]

Summary

• Introduces a first-class Project entity that groups channels and provides shared context to agents, eliminating the need to copy-paste canvas instructions into every new channel
• Projects are relay-stored Nostr events (kinds 50001-50003) with full CRUD: DB table, relay ingest/side-effects, REST API, SDK builders, and 5 MCP tools
• Channels gain an optional project_id FK — when set, the ACP harness automatically injects the project's prompt, repos, and environment into the [Context] block agents receive
• Includes 21 tests (SDK builders, ingest classification, DB CRUD), environment validation, empty-update guards
• Updates VISION.md and TESTING.md

What agents see automatically:

[Context]
Scope: channel
Channel: support-projects (#uuid)

[Project: sprout-dev]
Description: Sprout relay and desktop development
Environment: local
Repos: git@github.com:block/sprout.git

We're working on ~/dev/sprout here. Make sure we create new
worktrees for new work. Worktrees live in ~/.sprout/worktrees.

Files changed (24 files, +1105/-76)

• New: sprout-core/src/project.rs, sprout-db/src/project.rs, sprout-relay/src/api/projects.rs
• Core: kind.rs (3 new event kinds), schema.sql (projects table + FK)
• Relay: ingest.rs, side_effects.rs, router.rs, api/mod.rs
• SDK/MCP: builders.rs (3 builders + channel update), server.rs (5 tools), toolsets.rs
• ACP: pool.rs (project cache + fetch), queue.rs (format_prompt injection), relay.rs (ProjectInfo), main.rs
• Docs: VISION.md, TESTING.md

Future work

• Desktop UI (project management, project picker on channel creation)
• Project-level access control
• Cache TTL / invalidation

Test plan

• cargo check --workspace passes
• cargo test -p sprout-sdk -- project — 9 builder tests pass
• cargo test -p sprout-relay -- project — 6 ingest classification tests pass
• cargo test -p sprout-db -- project — 6 DB CRUD tests (require Postgres)
• cargo clippy --workspace -- -D warnings passes
• All pre-commit and pre-push hooks pass
• Manual: create project via MCP, create channel with project_id, verify agent receives project context

🤖 Generated with Claude Code

[wesbillman]

Design Status & Future Roadmap

Summary of concerns discussed during design and where each stands:

Concern	Status	Details
Stop copy-pasting canvas context	✅ Solved	Project prompt field auto-injected into agent [Context] block via ACP harness
Multiple repos per project	✅ Solved	repo_urls JSONB array on projects table
Mobile → create channel with project	✅ Solved	Project lives in relay, mobile can list/select, agents fetch from relay
Different local paths per user	🏗️ Architecture ready	Relay stores repo identity (git remote URL), local config resolves to disk paths. Need to build ~/.sprout/config.toml reader in ACP harness
Local agents vs blox agents in same channel	✅ Architecture right	environment field currently on project as a hint. Should evolve to per-agent-runtime config — same channel can have local + blox agents simultaneously
Worktree support	🏗️ Design ready	Add use_worktrees: bool to project schema, worktree root belongs in local config (~/.sprout/config.toml)
Agent routing (who picks up work)	📋 Future	Not project-specific — broader orchestration concern for when laptop is asleep but blox should handle work

Architecture: Relay = WHAT, Edge = WHERE

The project entity stores what agents work on (repo identities, shared instructions, preferences). Each agent runtime determines where and how (local paths, worktree roots, compute environment). This split means:

• Your laptop ACP reads the project from relay, merges with your local config (~/dev/sprout), agents see resolved paths
• Teammate's laptop ACP reads same project, merges with their local config (~/code/sprout), their agents see their paths
• Blox ACP reads same project, clones from git remote, agents work in blox workspace
• Mobile reads project metadata from relay for display/selection — never needs local paths

What's next (incremental, no rearchitecting needed)

1. ~/.sprout/config.toml — local config reader in ACP harness for repo path mapping
2. use_worktrees flag on project schema
3. Refine environment from project-level hint to agent-runtime config
4. Desktop UI — project management + project picker on channel creation
5. Project-level access control scoping