Long Context Dispatch

Metadata

• Trigger when: the current thread should stay in control but the analysis genuinely needs a much larger read-only context window.
• Do not use when: the task is ordinary coding, a short review, or anything that comfortably fits in the current thread.

Skill Purpose

Use the long-context worker as a read-only analysis surface, then synthesize the result back into the parent Codex thread without handing off final judgment.

Instructions

1. Gate the need honestly. If the task is not a strong long-context candidate, do nothing special and continue normally. If it is, get one short confirmation from the user before launch.
2. Launch the dispatcher with the exact script ~/.codex/skills/long-context-dispatch/scripts/long_context_dispatch.py. Put the request in a temporary Markdown file, pass an explicit --cwd, --reason, and --scope, and keep the child job read-only.
3. While the job runs, relay brief human updates instead of raw JSON. On success, read result.md and synthesize its findings back into the current thread. On failure, report the exit code, a short error summary, and the artifact paths with a concrete retry suggestion.

Non-Negotiable Acceptance Criteria

• The child worker stays read-only.
• One child job per parent request is the default; do not fragment the analysis without a reason.
• The parent Codex thread owns synthesis, editing decisions, and next steps.
• If the user wants an interactive long-context session rather than delegated analysis, route them to the dedicated launcher/app instead of forcing this skill.

Output

• A short launch note with job id and scope when dispatch starts.
• A synthesized result covering summary, key findings, cross-file connections, next actions, and open questions when the job succeeds.
• If the job fails: exit code, short error summary, and artifact paths.
• Next skill options (only if needed): $spec-bundle — turn large-scope analysis into implementation-ready artifacts; $adversarial-review — stress-test findings against actual changed behavior; $tool-scout — research the external landscape if the analysis exposed a buy-vs-build question.