harness-operator-brief

Harness Operator Brief

Produce one inline operator brief and one inline execution receipt. The managed policy controls communication. The released evaluator remains authoritative for identity, integrity, lifecycle state, and the canonical next action.

Required inputs

Require the exact skill name, one repository target, the structured launcher and expected identity for its exact external released evaluator, one requested outcome, declared non-effects, one supported source kind, one bounded UTF-8 source payload and digest, an ordered protected-content declaration, and any approved project terms.

Read the selected root's communication guide and this skill core. The new collection uses docs/engineering/harness/COMMUNICATION.md; older released roots select docs/engineering/TECHNICAL_COMMUNICATION.md. Use the installed root's choice and skill-contract.json for the requested output format.

For released-resources-v1, the entry comes from session activation and the communication guide from the selected evaluator's resources command. Use its returned path or content; do not require a repository copy. Missing selection or resources requires setup outside this read-only skill before the brief continues.

Plugin inputs and helper location

Use setup's already verified absolute environment Python and exact release identity. Its evaluator launcher is the actual argument array [ABSOLUTE_PYTHON, "-I", "-B", "-m", "se_harness"]; never look up harnessctl in the target or ambient PATH. Clear PYTHONPATH, put that environment's Scripts or bin first in the process PATH, and start direct calls outside the target. Missing inputs stop the brief; this skill does not run setup.

Use the helper from the selected plugin. Its content is not pinned in the repository lock. Read repository instructions again after compaction or a repository switch.

Pass the composed helper request as an absolute JSON file in host scratch space outside the target. Run the existing helper with separate arguments:

ABSOLUTE_PYTHON -I -B ABSOLUTE_PLUGIN_CORE/scripts/check_brief.py
  --request-json ABSOLUTE_REQUEST_JSON

The helper checks supplied source/output bytes and bindings. It does not select activation, query current state, write the brief, or construct the outer result and receipt. Follow the retained procedure below for those responsibilities. Keep both outputs inline and write no evidence into the target. The Codex metadata also disables implicit invocation; the explicit-name check applies regardless of host metadata support.

Procedure

1. Reject implicit or ambiguous activation and any explicit skill value other than harness-operator-brief. Accept only structured-harness-result or bounded-technical-text.
2. Run evaluator --version, released identity, and doctor with the supplied launcher. Require the expected exact identity and installed policy integrity.
3. If current repository or artifact state is needed and no current structured evaluator result was supplied, stop with current-state-result-required. Recommend harness-orient without invoking it. When current state is not needed, continue with the bounded supplied source; do not infer current state.
4. Validate the source digest, byte limits, protected span order, offsets, kinds, and digests before rendering. Treat every declared span as exact. Exclude semantically protected content from automatic paraphrase.
5. For an exact-output request, stop communication rendering and follow the existing exact-output path. If its governing procedure requires a canonical block alone, return that block unchanged, without result/receipt wrappers or a managed-profile claim; do not continue to steps 6–7. For eligible prose, compose one brief under operator-communication. Lead with the outcome or one required action. Preserve the accountable actor, conditions, qualifications, normative force, and one canonical next action.
6. Run scripts/check_brief.py over the source, rendered brief, output bindings, and changed-path observation. A binding uses UTF-8 byte offsets in the rendered brief and must identify the same bytes as its source span.
7. Return one se-harness-skill-result-v1 and one se-harness-execution-receipt-v1 inline. Include skill name and version, portable-core digest, policy path, selected profile, source digest, protected binding count, material deviations, residual uncertainty, evaluator identity, outcome, and zero changed paths.

Stop conditions

Stop before claiming a completed brief when activation, policy integrity, evaluator identity, source schema, source digest, profile selection, language, protected spans, output preservation, or zero-change evidence is missing, ambiguous, or invalid.

Stop and recommend harness-orient without invoking it when current state is required but no current structured evaluator result is supplied. Preserve an exact-output request unchanged. Ask for one accountable clarification when meaning cannot be preserved.

Boundaries

• Do not edit or retain files in the target.
• Do not transition artifacts, exercise a decision right, mutate Git, use credentials, access a network, release, deploy, or perform an external action.
• Do not download, search for, reproduce, or claim compliance with ASD-STE100.
• Do not expose source bodies in diagnostics or receipts, hidden reasoning, credentials, environment dumps, or host paths.
• Do not score grammar or claim that clarity proves technical correctness.
• Use the complete single-agent procedure. Do not spawn or coordinate workers.