A narrowly scoped community Codex skill that uses max-effort GPT-5.6 Luna subagents to absorb cold-context search, reading, extraction, and normalization while keeping reasoning, decisions, and implementation with the primary agent.
Route each stage by its work shape:
Deterministic transformation -> scripts or programmatic tool calling
Cold-context search and extraction -> one or more Luna-Max subagents
Deep reasoning and decisions -> primary agent
The skill can trigger implicitly without the user mentioning Luna or subagents when unfamiliar-context volume dominates the task. Luna locates, reads, extracts, labels, normalizes, and reports provenance. The primary agent forms hypotheses, resolves conflicts, judges implications, and writes the final answer.
- Web search, literature scouting, and source screening
- Bulk document, paper, and source-field extraction
- Codebase file, symbol, configuration, and dependency inventories
- Log, experiment, metric, and artifact collection
- Explicit-rubric labeling, deduplication, and data normalization
- Documentation, deprecation, and migration inventories
The skill intentionally does not delegate architecture, research direction, deep debugging, root-cause decisions, security assessment, final synthesis, high-stakes verification, or implementation.
Every delegate is spawned with:
model: gpt-5.6-luna
reasoning_effort: max
fork_turns: none
- A Codex environment with subagent or collaboration support
- Access to
gpt-5.6-luna - A spawn interface that supports model, reasoning-effort, and context-fork controls
If Luna or subagent tooling is unavailable, the skill instructs the primary agent to fall back to direct execution with the same task contracts.
Ask Codex:
$skill-installer install https://github.com/Ben-Lau1/delegate-work-to-luna/tree/main/skills/delegate-work-to-luna
Or use the system-provided skill-installer helper:
python3 "${CODEX_HOME:-$HOME/.codex}/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
--repo Ben-Lau1/delegate-work-to-luna \
--path skills/delegate-work-to-lunaThe skill will be available on the next Codex turn.
These examples follow the latest main. For a reproducible install, replace main in the URL or pass --ref <release-tag> to the helper.
The installer intentionally refuses to overwrite an existing skill. To upgrade safely, move the current installation aside, install the new copy, and keep the backup until the new version has been verified:
luna_skills_dir="${CODEX_HOME:-$HOME/.codex}/skills"
mv "$luna_skills_dir/delegate-work-to-luna" "$luna_skills_dir/delegate-work-to-luna.backup"
python3 "$luna_skills_dir/.system/skill-installer/scripts/install-skill-from-github.py" \
--repo Ben-Lau1/delegate-work-to-luna \
--path skills/delegate-work-to-lunaStart a new Codex turn before verifying the updated skill. If installation fails and no new destination was created, restore the backup by moving it back to delegate-work-to-luna.
To disable the skill without deleting it, move its directory outside $CODEX_HOME/skills. Move it back and start a new Codex turn to enable it again.
Explicit invocation:
$delegate-work-to-luna search official sources and extract the current API requirements with direct links
$delegate-work-to-luna scan these experiment directories and extract configs, metrics, missing fields, and provenance
$delegate-work-to-luna inventory every deprecated API occurrence in this repository with paths and line numbers
Implicit invocation is enabled explicitly. Natural requests such as these should match without naming the skill:
Screen these papers and return titles, authors, methods, datasets, metrics, and DOI links.
Search the official documentation and collect the requirements, defaults, and version constraints.
Scan these logs, group repeated messages, and record timestamps, counts, and first occurrences.
Requests such as “design the architecture,” “find the root cause and patch it,” or “perform a security review” should not trigger this skill. The primary agent may still delegate a separable evidence-collection stage, but Luna must not make the judgment. Deterministic bulk transformations should use scripts.
- The primary agent retains user intent, architecture decisions, integrated validation, and consequential actions.
- Delegate output is treated as an untrusted evidence bundle until verified.
- Luna lanes are read-only: no code or file edits, external writes, deployments, destructive actions, or production mutations.
- Luna reports observations and provenance without issuing final recommendations or conclusions.
- Recursive agent trees and duplicate work are prohibited by default.
The skill passes the Codex skill validator. In a fresh positive forward test, a natural request to inventory fields across up to 100 repository files implicitly launched a read-only luna_inventory lane. In a fresh negative test, a global payment-architecture decision remained with the primary test agent and launched no Luna lane. Both runs were stopped after routing behavior was observed, so these tests validate selection boundaries rather than result quality or completeness.
This is an early community release. Feedback and reproducible task traces are welcome.
This project is not affiliated with or endorsed by OpenAI.