Research Harness

Use the external project as authority. nullius owns lifecycle, approvals and state; the host owns its agent loop, tools and session compaction. This skill routes work and records results, rather than replacing either runtime.

Recover the current task

1. Work in the external project root. If .nullius/HARNESS exists, first obtain ./.nullius/bin/nullius status --json; use nullius status --json if the local launcher is unavailable. A missing/broken handshake or launcher is repaired with nullius init --runtime-only, followed by the local status call.
2. Stop on a status error or failed machine gate. Carry out a reported repair and continue; advisory warnings become recorded work, not automatic blocks. If any A1-A5 approval is pending, stop at that boundary. Silence is not approval.
3. Read research_plan.md#Current Status, then use project_index.md to locate the current adopted result and the original files relevant to this task. Read applicable charter/contract constraints and relevant notebook sections. Reacquire complete source whenever a claim or verification depends on it.
4. Before restart, context reset or handoff, save required inputs, choices, failures, unknowns, next action and acceptance command in durable project material. Process-local stores are not checkpoints. A resumed session ID does not establish that needed bytes survived.

Read recovery details for anchor freshness, status classifications, missing state, initialization or migration. For task-scoped context and phase changes, use task briefs and handoff.

Route only the work that applies

Ordinary bounded edits stay with the current owner. Use research-team for a research milestone requiring execution and independent verification. Keep lifecycle/status/approval/export operations on nullius; provider MCP tools are capabilities, not a second project authority. For an unscoped question, nullius workflow-plan --recipe research_brainstorm records planning only.

An unavailable required skill leaves the result unverified. Tooling fixtures and ordinary code tests do not become scientific results. Review identity, method and input independence must be assessed separately. Read routing details when deciding the scope or the required verification for a result.

Load operational detail when triggered

Durable conclusions

Engine and file mode owe the same result-specific evidence. Fold verified reasoning into the human notebook, contract and immutable report with source pointers. Structural links, status, hashes, receipts and reviewer agreement do not by themselves prove scientific truth. Closeout requires the current unscoped notebook structure check and accepted full-text content review; selected-claim or delta reviews cannot replace it. Keep complete original material available to verification even when the working context is brief. Walk research-integrity M1–M7 before folding or requesting final approval. Do not treat missing, partial, failed, unavailable, invalid or stale evidence as accepted. Read the closeout reference before performing these operations.