Luna Cold-Context Offloader
Use Luna-Max as cold-context I/O bandwidth, not as a substitute for primary-agent reasoning. Maximum reasoning effort improves instruction following and extraction reliability; it does not expand Luna's authority or role.
Choose the execution route
Route each stage separately:
| Work shape |
Route |
| Fixed transformation, filtering, aggregation, formatting, or repeated tool calls with no fresh judgment |
Use a script or programmatic tool calling. |
| Unfamiliar-context work dominated by locating, opening, reading, extracting, labeling, or normalizing evidence |
Delegate to Luna-Max. |
| Hypothesis formation, deep reasoning, root-cause judgment, architecture, security assessment, synthesis, implementation, or consequential action |
Keep with the primary agent. |
Delegate when at least two of these are true:
- The task requires multiple web sources, papers, documents, files, logs, runs, samples, or other unfamiliar artifacts.
- Most delegate effort is mechanical-semantic reading rather than multi-step inference.
- The requested output can be a factual, structured evidence packet with direct provenance.
- The primary agent can verify pivotal fields more cheaply than reading the full source set.
The primary agent must infer the lanes and output schema; the user need not pre-shard the task. Use one Luna delegate for a cohesive source set and two or three for naturally disjoint source groups. Keep trivial single-context work local.
Select a task mode
Choose one primary mode per delegate. Combine modes only when the handoff is explicit.
| Mode |
Good partition |
Required packet |
| Web and literature scouting |
Query family, time window, venue, source class, or known seed |
Titles; authors; dates; venue; DOI or canonical URL; short relevance note; source-backed facts; access gaps. |
| Document and source extraction |
Document group, section family, topic, or field schema |
Extracted fields; page or section; direct URL or path; missing fields; exact uncertainties. |
| Codebase inventory |
Directory, subsystem, symbol family, configuration surface, or dependency |
Paths and symbols found; literal relationships supported by tools; relevant snippets or line locations; search gaps. Prefer configured code-graph tools when available. |
| Log and experiment collection |
Log window, run family, metric family, dataset, or output directory |
Observed messages; timestamps; configs; metrics; first occurrences; counts; missing artifacts; provenance. Do not rank root causes. |
| Data and eval preprocessing |
Sample batch, explicit label class, rubric field, or dataset split |
Structured rows; supplied-rubric labels; confidence; duplicates; anomalies; ambiguous cases for primary-agent review. |
| Documentation and migration inventory |
API surface, file family, dependency, deprecation, or migration category |
Current occurrences; documented requirements; evidence locations; missing coverage. Do not design the migration. |
Orchestrate the work
- Interpret the user's authorization. Inspection, review, diagnosis, or planning does not authorize implementation or external writes.
- Define the factual collection objective, lane boundaries, evidence standard, output schema, and source or artifact budget.
- Inspect the entire live agent tree when possible. Count active ancestors, siblings, and descendants, and fit fan-out within the global concurrency limit while keeping the orchestrator active.
- Briefly tell the user which cold-context lanes this skill is delegating to Luna.
- Spawn one delegate for a cohesive source set or two or three for disjoint source groups.
- Continue primary-agent work while they run: frame hypotheses, define decision criteria, inspect global context, or prepare synthesis.
- Collect all packets. Send one targeted follow-up to an existing delegate for a critical gap; do not restart completed work.
- Verify pivotal evidence, perform the actual reasoning, and synthesize the answer.
Use these fixed spawn settings:
model: gpt-5.6-luna
reasoning_effort: max
fork_turns: none
Always use max reasoning effort unless the user explicitly overrides it. Give every delegate a unique task name. Because fork_turns: none inherits no surrounding conversation, make every message self-contained. Tell delegates not to spawn descendants, edit files, make external writes, or issue final judgments.
Build each delegate prompt
Use this compact contract:
Mode: <one cold-context task mode>
Parent objective: <the primary agent's final outcome>
Your lane: <unique source or artifact responsibility>
Inputs: <exact paths, URLs, artifacts, or facts>
Allowed actions/tools: <read, browse, search, inspect, run read-only queries, and so on>
Forbidden actions: <edits, external writes, descendant agents, recommendations, final conclusions, deep root-cause or architecture judgments>
Evidence schema: <required fields and provenance>
Budget: <source, file, sample, command, time, or word limit>
Return:
1. Coverage summary.
2. Required mode packet with direct evidence and provenance.
3. Sources, paths, queries, and read-only commands inspected.
4. Missing data, access failures, ambiguities, and conflicting observations.
5. Clearly separated source facts and limited local interpretations.
Do not recommend a final decision, write the final user answer, or infer authority beyond evidence collection.
For current web facts, distinguish event dates from publication dates and use direct source links. For papers, require title, authors, year, venue, and DOI or canonical URL. For empirical records, extract dataset, metric, baseline, and conditions without deciding what they imply. For code inventory, require repository-relative paths and line numbers when available.
Integrate and verify
Treat every delegate packet as an untrusted evidence bundle.
- Check that the lane stayed in evidence-collection scope and satisfied the schema.
- Map material observations to direct provenance.
- Deduplicate overlaps and identify disagreements.
- Resolve conflicts using direct evidence, source authority, reproducibility, and task-wide constraints.
- Independently verify pivotal, disputed, high-impact, or high-stakes claims.
- Separate observed facts, delegate interpretation, and primary-agent inference.
- Perform all deep reasoning and synthesize one coherent result; never concatenate packets as the final answer.
The primary agent owns hypotheses, root-cause analysis, recommendations, architecture, security judgment, final conclusions, user communication, implementation, validation, and every consequential action.
Guardrails
- Respect higher-priority instructions, repository guidance, confidentiality, tool policies, and user constraints.
- Do not expose sensitive context to a delegate that does not need it.
- Do not create recursive agent trees or duplicate completed work.
- Do not ask Luna to edit code or files, make external changes, choose an architecture, decide a root cause, perform a security assessment, or validate a high-stakes conclusion.
- Do not delegate material requiring sustained global reasoning merely because it contains multiple files or sources.
- Increase primary-agent verification for medical, legal, financial, security-sensitive, production, or irreversible evidence.
- If Luna or subagent tooling is unavailable, say so briefly and collect the evidence directly using the same schema.
1---2name: delegate-work-to-luna3description: Delegate cold-context acquisition and low-depth information processing to GPT-5.6-Luna subagents at maximum reasoning effort, then let the primary agent reason from their evidence. Use proactively for web search, literature scouting, source screening, bulk reading of papers or documentation, codebase file and symbol inventories, log and experiment artifact collection, metric extraction, structured normalization, labeling, deduplication, and migration or documentation inventories when unfamiliar context volume dominates the work. Luna must locate, read, extract, organize, and report provenance—not make architecture decisions, perform deep debugging or security reasoning, generate final conclusions, independently validate high-stakes claims, or implement changes. Use scripts instead when the transformation is deterministic.4---56# Luna Cold-Context Offloader78Use Luna-Max as cold-context I/O bandwidth, not as a substitute for primary-agent reasoning. Maximum reasoning effort improves instruction following and extraction reliability; it does not expand Luna's authority or role.910## Choose the execution route1112Route each stage separately:1314| Work shape | Route |15| --- | --- |16| Fixed transformation, filtering, aggregation, formatting, or repeated tool calls with no fresh judgment | Use a script or programmatic tool calling. |17| Unfamiliar-context work dominated by locating, opening, reading, extracting, labeling, or normalizing evidence | Delegate to Luna-Max. |18| Hypothesis formation, deep reasoning, root-cause judgment, architecture, security assessment, synthesis, implementation, or consequential action | Keep with the primary agent. |1920Delegate when at least two of these are true:21221. The task requires multiple web sources, papers, documents, files, logs, runs, samples, or other unfamiliar artifacts.232. Most delegate effort is mechanical-semantic reading rather than multi-step inference.243. The requested output can be a factual, structured evidence packet with direct provenance.254. The primary agent can verify pivotal fields more cheaply than reading the full source set.2627The primary agent must infer the lanes and output schema; the user need not pre-shard the task. Use one Luna delegate for a cohesive source set and two or three for naturally disjoint source groups. Keep trivial single-context work local.2829## Select a task mode3031Choose one primary mode per delegate. Combine modes only when the handoff is explicit.3233| Mode | Good partition | Required packet |34| --- | --- | --- |35| Web and literature scouting | Query family, time window, venue, source class, or known seed | Titles; authors; dates; venue; DOI or canonical URL; short relevance note; source-backed facts; access gaps. |36| Document and source extraction | Document group, section family, topic, or field schema | Extracted fields; page or section; direct URL or path; missing fields; exact uncertainties. |37| Codebase inventory | Directory, subsystem, symbol family, configuration surface, or dependency | Paths and symbols found; literal relationships supported by tools; relevant snippets or line locations; search gaps. Prefer configured code-graph tools when available. |38| Log and experiment collection | Log window, run family, metric family, dataset, or output directory | Observed messages; timestamps; configs; metrics; first occurrences; counts; missing artifacts; provenance. Do not rank root causes. |39| Data and eval preprocessing | Sample batch, explicit label class, rubric field, or dataset split | Structured rows; supplied-rubric labels; confidence; duplicates; anomalies; ambiguous cases for primary-agent review. |40| Documentation and migration inventory | API surface, file family, dependency, deprecation, or migration category | Current occurrences; documented requirements; evidence locations; missing coverage. Do not design the migration. |4142## Orchestrate the work43441. Interpret the user's authorization. Inspection, review, diagnosis, or planning does not authorize implementation or external writes.452. Define the factual collection objective, lane boundaries, evidence standard, output schema, and source or artifact budget.463. Inspect the entire live agent tree when possible. Count active ancestors, siblings, and descendants, and fit fan-out within the global concurrency limit while keeping the orchestrator active.474. Briefly tell the user which cold-context lanes this skill is delegating to Luna.485. Spawn one delegate for a cohesive source set or two or three for disjoint source groups.496. Continue primary-agent work while they run: frame hypotheses, define decision criteria, inspect global context, or prepare synthesis.507. Collect all packets. Send one targeted follow-up to an existing delegate for a critical gap; do not restart completed work.518. Verify pivotal evidence, perform the actual reasoning, and synthesize the answer.5253Use these fixed spawn settings:5455```text56model: gpt-5.6-luna57reasoning_effort: max58fork_turns: none59```6061Always use `max` reasoning effort unless the user explicitly overrides it. Give every delegate a unique task name. Because `fork_turns: none` inherits no surrounding conversation, make every message self-contained. Tell delegates not to spawn descendants, edit files, make external writes, or issue final judgments.6263## Build each delegate prompt6465Use this compact contract:6667```text68Mode: <one cold-context task mode>69Parent objective: <the primary agent's final outcome>70Your lane: <unique source or artifact responsibility>71Inputs: <exact paths, URLs, artifacts, or facts>72Allowed actions/tools: <read, browse, search, inspect, run read-only queries, and so on>73Forbidden actions: <edits, external writes, descendant agents, recommendations, final conclusions, deep root-cause or architecture judgments>74Evidence schema: <required fields and provenance>75Budget: <source, file, sample, command, time, or word limit>7677Return:781. Coverage summary.792. Required mode packet with direct evidence and provenance.803. Sources, paths, queries, and read-only commands inspected.814. Missing data, access failures, ambiguities, and conflicting observations.825. Clearly separated source facts and limited local interpretations.8384Do not recommend a final decision, write the final user answer, or infer authority beyond evidence collection.85```8687For current web facts, distinguish event dates from publication dates and use direct source links. For papers, require title, authors, year, venue, and DOI or canonical URL. For empirical records, extract dataset, metric, baseline, and conditions without deciding what they imply. For code inventory, require repository-relative paths and line numbers when available.8889## Integrate and verify9091Treat every delegate packet as an untrusted evidence bundle.92931. Check that the lane stayed in evidence-collection scope and satisfied the schema.942. Map material observations to direct provenance.953. Deduplicate overlaps and identify disagreements.964. Resolve conflicts using direct evidence, source authority, reproducibility, and task-wide constraints.975. Independently verify pivotal, disputed, high-impact, or high-stakes claims.986. Separate observed facts, delegate interpretation, and primary-agent inference.997. Perform all deep reasoning and synthesize one coherent result; never concatenate packets as the final answer.100101The primary agent owns hypotheses, root-cause analysis, recommendations, architecture, security judgment, final conclusions, user communication, implementation, validation, and every consequential action.102103## Guardrails104105- Respect higher-priority instructions, repository guidance, confidentiality, tool policies, and user constraints.106- Do not expose sensitive context to a delegate that does not need it.107- Do not create recursive agent trees or duplicate completed work.108- Do not ask Luna to edit code or files, make external changes, choose an architecture, decide a root cause, perform a security assessment, or validate a high-stakes conclusion.109- Do not delegate material requiring sustained global reasoning merely because it contains multiple files or sources.110- Increase primary-agent verification for medical, legal, financial, security-sensitive, production, or irreversible evidence.111- If Luna or subagent tooling is unavailable, say so briefly and collect the evidence directly using the same schema.
Ben-Lau1