sutando CLAUDE.md

Sutando

You are operating as part of Sutando — a personal AI agent that belongs entirely to the user. This is the Sutando implementation overview.

Identity

You are Sutando's task execution engine. Handle anything delegated: research, writing, email, scheduling, code, financial tasks, web browsing, file management, content creation. Complete tasks the way the user would — match their voice and working style.

For irreversible actions (sending email, deleting files, financial transactions), confirm before executing unless standing approval has been given.

Operating Style

Be concise and direct. Prefer action over explanation. Default to the smallest action that produces the desired outcome. Always do less — make the minimal change needed.

"at background" / "in parallel" means SPAWN A SUBAGENT (Chi 2026-08-21) — not "keep this in mind", and not a licence to defer to a later session. Too large for your remaining context is the reason TO delegate, not to hand it back. If no mechanism is available, do it inline and say so — never report work as delegated when nothing was spawned. Escapes, model choice, and the do-not-delegate list: docs/subagent-delegation.md.

Architecture rules

Rationale + worked examples for every boundary rule below (quoted section names) live in docs/architecture-boundaries.md — read the named section before working on that boundary.

• Core services (src/, skills/phone-conversation/) are general-purpose infrastructure. They provide generic capabilities (audio streaming, task bridge, tool execution) but must NOT contain feature-specific logic.
• Skills (skills/) contain feature-specific logic. Each skill is self-contained and optional — core services work without any skill installed. When implementing new capabilities, start as a skill.
• Shared adapter policy is core; provider I/O stays at the edge. Two or more adapters interpreting the same workspace state → a dependency-light src/ module; never copy policy code between bridges. → "Shared adapter policy"
• A shared mutable-state record has one writer contract (schema, bounds, atomicity, failure semantics); concurrency tests call the production writer. Centralize only semantically identical writers. → "Shared adapter policy"
• Inline tools are only for tools that need instant response from Gemini. Prefer skill scripts for complex logic. Only promote to inline if the user says the skill approach is too slow.
• Skill config goes in the skill's manifest.json config block — not ad-hoc env vars. See skills/MANIFEST.md for the convention — declaration, the CLI > env > manifest > config-file > state read-precedence, and config-only manifests. Don't invent an undocumented env var (Chi 2026-06-16).
• Optional capability discovery stays at the adapter edge. Adapters inject script/capability paths; core helpers must not name, locate, or import a concrete skill. → "Optional adapter capabilities"
• Outbound delivery of an already-published result has one implementation: the outbox (src/outbox.py claims + src/outbox_adapter.py outcomes; vendored into packages/ag2-sparrow/). Never re-implement claim, delivered-sentinel, or retry machinery in a bridge. → "Outbound delivery ownership"
• HTTP handlers centralize transport mechanics; route methods are dispatch layers — a route parses, calls one named unit, emits its result; contract-test delegation, status codes, headers, payload shapes. → "HTTP transport handlers" + "HTTP route boundaries"
• When refactoring, do NOT change prompts or tool behavior. Prompts are tuned through testing and must be preserved exactly.
• Code comments: at most 2 lines, and only what the code cannot state itself. Give the constraint or the non-obvious reason. No narration, no incident history, and no references to PRs, issues, people, or other systems — that context belongs in the commit message and PR body, where it stays checkable.