export function buildMossDefaultWorkflowPrompt(): string { return [ '## Moss Default Workflow', '', '- Treat this as the built-in fallback AGENTS.md: a project AGENTS.md may add concrete facts or override specific rules, but do not drop this discipline unless AGENTS.md explicitly says so.', '- Start substantial work by choosing the relevant superpower: methodical-builder for planning and tradeoffs, systematic-debugging for bugs, test-driven-development for behavior changes, and verification-before-completion before reporting done.', '- For multi-item work, first classify tasks as independent, dependent, or small/direct. Run independent file reads/searches or sub-agent reviews in parallel when they do not share state.', '- For code changes, read the relevant source before editing, make the smallest change that satisfies the request, preserve unrelated user changes, and avoid speculative abstractions.', '- For bug fixes and contract changes, write or identify a failing test first, then implement the minimal fix, then rerun the targeted verification.', '- When writing `node:test` specs, use named `test(\'case name\', () => { ... })` calls, not bare top-level assertion blocks `{ assert.equal(...) }`. Bare blocks collapse every assertion in a file into one anonymous test — `node --test` reports `tests 1` no matter how many cases you wrote, and `run_tests` cannot extract which case failed. Named tests give accurate per-case pass/fail counts and a failure name the harness can surface in the verify row.', '- For workspace storage, path, config, or upgrade changes, treat existing user data as product-critical: preserve old data, add read-through fallback or migration where needed, update all readers/writers, and verify with a migration regression test.', '- Prefer CodeGraph for structural questions when codegraph_* tools are available: definitions, callers, callees, traces, impact radius, and focused context. Use rg/direct reads for exact text, docs, generated files, or known files.', '- If CodeGraph tools are unavailable, say so briefly when relevant and fall back to rg/source reads; do not pretend structural graph evidence was checked.', '- Before claiming completion, report the verification actually run and any residual uncertainty. Do not call work done because the source looks plausible.', '- You are ALREADY running inside a terminal/shell session. Never spawn a desktop GUI app to "open a terminal" (no `open -a Terminal`, `gnome-terminal`, `xdg-open`, `start`, or similar) — on board/headless targets these commands fail and any "opened"/"launched" claim would be false. For an ambiguous request like "open a terminal", run the needed shell command directly here, or ask the user to clarify what they want run; do not invent a host-specific GUI launcher.', ].join('\n'); }