interactive-terminal-mcp AGENTS.md

Agent Guidelines

Scope discipline (CRITICAL)

• Break every task into the smallest possible independent scope. One agent = one concern.
• Fan out: launch multiple scoped agents in parallel when work is independent.
• Sequential when dependent: scout → plan → dev → review. Never dev before scout reports.
• Never do by hand what an agent can do. Manual edits are error-prone and waste tokens.
• Keep agent prompts focused — one file or one feature per agent. Large multi-file rewrites in a single agent prompt time out or produce incomplete work.

Workflow

1. Scout (parallel, read-only): investigate the codebase, report findings. 3-5 agents fanned across different areas.
2. Consolidate: merge scout findings into a prioritized plan.
3. Dev (sequential, one concern per agent): implement fixes one scope at a time. Build + test after each.
4. Review (read-only): verify the dev agent's work. Report GREEN or findings.
5. Fix (dev): address review findings.
6. Repeat review → fix until GREEN.

Anti-patterns to avoid

• Launching scout + dev simultaneously (dev needs scout findings first).
• One giant agent prompt covering multiple files/concerns.
• Manual hand-editing when agents are available.
• Asking the user questions when the task is clear — proceed autonomously.
• Implementing backwards compatibility the user didn't ask for — confirm first.

Project-specific notes

• The daemon is load-bearing. Sessions must outlive any single MCP process, so internal/daemon owns every PTY and everything else is a client of it. Never move terminal state into the MCP server or the CLI.
• Run -race on anything touching sessions. The daemon multiplexes PTY readers, emulator writes, log writers, broadcasters, and socket handlers. Two real bugs (a lost-output drain and an emulator deadlock) were caught only under the race detector.
• The emulator's reply pipe must always be drained. Programs query the terminal on startup and block for the answer; an undrained emulator deadlocks every session. See Session.answerQueries.
• Alternate-screen output never reaches the transcript. That is correct terminal behaviour, not a bug. it_tail compensates by returning the live screen; don't "fix" it by writing alt-screen frames to the log.
• Agent-facing strings are part of the contract. Every error carries a concrete next tool call, and truncation always reports what was dropped plus a path to the whole file. Tests in internal/mcpserver pin this.