Triage and archive accumulated Codex conversations without losing wanted work. Use when cleaning up old Codex threads, deciding which conversations still need attention, or resuming a large conversation cleanup. Do not use for ordinary backlog capture, feature implementation, or issue-tracker cleanup.
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Clear tasks that no longer need attention while giving every still-wanted outcome a durable continuation. A task and an outcome have separate lifecycles: one outcome may span several tasks, and one task may contain several outcomes. Archiving is reversible; it does not delete the task.
For each task in an archive proposal, ask: can the user stop paying attention to it without losing work they still want? The answer must be clear without opening the source conversation. Keep the fuller inventory and supporting evidence in the cleanup checkpoint, where the next pass can resume it.
Establish the repository or project scope. Use Codex thread tools to list and read tasks, and to archive and verify them. If those tools are unavailable, report the limit rather than claiming to have inspected or archived tasks. Treat titles, age, and summaries as discovery hints; read recent turns and relevant completion evidence. Earlier promises may require reading farther back. Check current code, decisions, specs, and backlog before treating a claimed implementation or destination as complete. An ADR alone does not prove implementation.
Inventory IDs before archiving so pagination shifts cannot skip candidates. Work in bounded batches and keep a checkpoint in the cleanup task or a linked working note: scope, inventory progress, reviewed IDs and reasons, preservation destinations, approved actions, verified archives, and unresolved cases. This checkpoint tracks cleanup progress, not product backlog. On resumption, recheck tasks with new activity and retry unconfirmed operations; do not reread an unchanged completed batch. Keep the cleanup task and running work active.
For a large inventory across products or long histories, subagents may read bounded sets of exact IDs and return compact evidence about completion, remaining user intent, existing destinations, and uncertainty. Give them read-only work. The main agent verifies their findings against current task and repo state, integrates the proposal, and owns preservation and archiving. A small batch needs no delegation.
First decide whether work must continue in this task. Keep active implementation and conversations the user expects to resume there. A dirty worktree alone does not keep a task active; archiving never authorizes discarding its worktree, edits, or branch. Git status informs preservation; it does not decide whether the conversation still needs attention.
For a task that no longer needs attention, account for every remaining outcome:
A blocked task can close when its wanted outcome has a destination and the user does not need to resume that task. Pinning, historical importance, and valuable reasoning do not by themselves require attention. When another active task owns the work, name it as the continuation. Completed delegations and reviews can close once their results and any remaining integration work are accounted for.
Lead with the scope and recommended batch in ordinary language. When confusion is possible, say that archiving changes the Codex task list, not the files or worktree. For each proposed archive, explain what the task settled and where any still-wanted work will continue. For a task that must stay active or has unclear intent, explain the specific reason and the decision that would change it. Give only the detail needed to test those claims without opening the source task. A simple closure may take one sentence; a task with several outcomes may need more. Group tasks only when they truly share a reason. Keep inventories, counts, and raw evidence in the checkpoint rather than making the user read an audit report. Ask only questions whose answers change the proposed action.
After the readable judgments, show the exact archive IDs and proposed backlog edits as execution details before requesting approval when approval is required. An audit request alone is not permission to archive. Honor the user's standing approval preferences. Approval for one batch does not authorize the next batch.
Once authorized, recheck each authorized task for new activity. Preserve authorized outcomes first and confirm the write succeeded. Archive only authorized IDs with Codex thread tools, then verify each in archived-task state. If preservation fails, leave its source task unarchived. Do not infer success from a mutation response alone.
Report reviewed, archived, kept active, unresolved, failed, and not-yet-reviewed counts separately. Report backlog changes separately. Keep exact IDs, titles, reasons, destinations, and operation results in the checkpoint so the next batch starts from verified state.
EpicenterHQ/epicenter
7cc869b
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.