m1nd-operator

m1nd Operator

This is the deep execution manual that complements the short m1nd-first doctrine.

Use this skill when m1nd should be the first layer of truth for the task.

m1nd is strongest when structure, connected context, blast radius, continuity, docs/code bindings, or multi-agent coordination are the bottleneck. It is not the replacement for rg, the compiler, runtime logs, or the test runner.

Default Stance

Default to m1nd before grep, glob, or manual file reads.

The first question is not "which shell command should I run?" It is "can m1nd answer or narrow this directly from graph structure, connected context, docs, or cross-domain bindings?"

Only skip the m1nd first pass when:

• the user already gave the exact file and exact lines
• the question is pure compiler/test/runtime truth
• the task is a trivial local file action with no search or structural uncertainty

Trust-first is calibrated-first, not blind-first. Default to m1nd, AND hold its honest ruler in the same hand: it narrows and connects, never replaces the compiler/tests/runtime; and it still has live stumbles measured in the field — federated seek/federate can hang past 60s (bound-timeout it or fall back), a freshly-memorized claim can lag its own semantic recall for a tick (the write is durable — re-seek or widen, don't conclude absence), and the runner naming lane can fall back to heuristics under load. In each case WORK AROUND + one field-report line; never blind insistence. The reason to still go first: the calibrated absence (abstain/caller_root_mismatch/gathering) is itself the answer that most often changes the decision — the counterfactual test for attribution.

Trained Agent Loop

For unfamiliar repo work, audits, bug hunts, reviews, and risky changes, the measured high-signal pattern starts by never starting cold:

1. Call north(task) FIRST, before reading or editing anything. It is the in-session front door. One round-trip returns binding trust (trust_mode; repair travels with it when degraded), task context (focus nodes + PageRank anchors), prior cross-session memory (each claim with real age + author — absent, never faked, when unknown), a sufficiency signal, one next_move, and honest_gaps (what m1nd does NOT know) — plus code, the real SOURCE of the top focus nodes (the symbol's own lines, or the file head when a node names no symbol) for up to 3 files inside a ~2,000-char budget, each slice declaring total_lines vs lines_returned and truncated. Read that payload instead of grepping or re-opening what north just named; widen the first call with code_budget_chars rather than paying a second round-trip, and pass code: false when you want orientation only. north composes trust_selftest + orient + boot_memory + focus + surgical_context/batch_view; reach for the pieces directly only when you need just one. Heed reception: reception.match == "caller_root_mismatch" means the bound graph does NOT cover your current repo — do not trust retrieval for it; read reception.options[]. Generic cross-root ingest remains withdrawn (project_root absent from the published schema), and over the wire brain.bootstrap.birth answers every client with human_gesture_required. The way forward for a brainless repo is the human's one-time ceremony m1nd init --birth <repo>: offer that command and stop. Until the human runs it, continue only against the bound graph with the mismatch warning intact, or reconnect to an owner that already hosts the intended repo. Absent/null = your root matches the brain serving you. Reception governs WRITES too, not only reads: a read under mismatch is a warning, a WRITE under mismatch is prohibited (see Write-Mode Laws below) — and no public bootstrap lifts it. VERSION SKEW: this skill describes the CURRENT product; check binding.fingerprint.binary_version (and its binary_drift warning) in the same packet before trusting either era's rule. A legacy 1.4.x served owner still HAS the one-call bootstrap, and there ingest {project_root: <your repo root>} is the correct foreign-repo gesture — it mints a SEPARATE per-project brain, never replacing the bound one. A plain ingest {path} from a foreign root is wrong in EVERY era: on a legacy owner it wholesale-REPLACES the bound brain (2026-07-24 incident — a 29k-node brain swapped for a foreign graph, restored the same day).
2. If north returns needs_ingest (empty/unbound graph), re-scan and call north again. needs_ingest is a REAL answer, not a failure. On a brain that already DECLARES your root, the repair is yours to run and needs no human: ingest {mode:"refresh", path: <your root>} from exactly that root (Law 3 below). On a root no brain declares, it is not a refresh — no agent mints a brain on any transport — and the honest move is to OFFER the human their one-time ceremony, m1nd init --birth <repo>, and stop (or reconnect to an owner that already hosts the repo). A plain ingest {path} (mode replace) remains refused, and its refusal now names that same command.
3. Act on verdicts, do not override them (see below).
4. Prove final truth with direct source reads, tests, compiler/runtime output, and focused probes.
5. Before edits/reviews, run impact, validate_plan, and usually surgical_context_v2.
6. Record the investigation path: m1nd calls, recovery decisions, files inspected, commands run, and fallback reason.