sweep-opp-runs

Per-opp retention prune, not orphan sweep. For every opp folder under ACE/, sort its runs/<run-id>/ children by createdTime descending, keep the newest N, propose the rest as trash candidates. Human approves, atom executes via drive_trash_file (reversible — 30-day Drive bin).

This is purely about retention to control /ace:status and other listing load times. It does not consult the live-set; orphan opp folders are sweep-drive's job. opp-runs assumes every opp under ACE/ is legitimate and only trims its run history.

Inputs

• keep (required, int ≥ 1) — the number of newest runs to retain per opp. Sweep orchestrator surfaces this from the user; default is 3 if the user accepts the prompt without changing it.
• sweepFolder — the timestamped sweep folder built by the orchestrator (e.g. ACE/_sweep/<timestamp>/); products land here.
• mode — recommend (default) | execute. In recommend the skill is report-only: build + render the per-opp prune plan + return it; no trashing. In execute it trashes only approvedIds. The human-confirmation gate between the two is the orchestrator's job — see agents/sweep.md § Human-confirmation gate.
• approvedIds (execute mode only) — the exact run-folder ids the human approved in chat. Trash nothing outside this set.

There is no live-set dependency. If the orchestrator built one (because the user ran a different system first in the same /ace:sweep all-style invocation), ignore it.

Products

• ACE/_sweep/<timestamp>/opp-runs-prune.md — human-readable plan grouped by opp.
• ACE/_sweep/<timestamp>/opp-runs-prune.yaml — machine-readable per-opp plan (for replay / partial approval).
• For each approved run folder: a drive_trash_file op.

Process

1. Enumerate opps. drive_list_folder on ACE_DRIVE_ROOT_FOLDER_ID. Filter to folders only (mimeType == application/vnd.google-apps.folder). Skip names starting with _ (e.g. _sweep/, _archive/, ai-input-creation-runs/ is NOT a sweep target — it doesn't start with _, but it's also not an opp; the runs-folder probe in step 3 naturally skips it).
2. Locate runs/ per opp. For each opp folder, drive_list_folder on the opp folder; find the child named exactly runs (folder mimeType). If none exists, skip the opp (nothing to prune).
3. List runs. drive_list_folder on the runs/ folder. Filter to folders. These are individual run dirs (typically YYYYMMDD-HHMMSS-<short> or similar; do not parse the name — sort by createdTime from the Drive metadata, not the slug).
4. Sort + slice. Sort runs by createdTime descending (newest first). The first keep runs are retained. The remainder are prune candidates.
5. Build the per-opp plan. For each opp with >0 candidates, collect:
   • oppSlug — the opp folder name
   • kept[] — the keep newest runs (id, name, createdTime)
   • candidates[] — the runs to trash (id, name, createdTime)
6. Render the markdown report. Per-opp section with two short tables (kept vs candidates). Headline at the top: "Prune plan — keep newest N per opp; M opps affected; K runs to trash." Write to ACE/_sweep/<timestamp>/opp-runs-prune.md and the YAML peer.
7. Recommend — stop here in mode: recommend. Return the per-opp prune plan (candidates to trash, grouped by opp — each with id, name, createdTime; reversibility: 30-day Drive bin) + the report Drive link, to the orchestrator. Trash nothing. Do not try to prompt the human from this skill — a dispatched subagent can't reach them; the orchestrator runs the confirmation gate (see agents/sweep.md § Human-confirmation gate).

Execute phase (mode: execute only)

Runs only when the orchestrator re-dispatches with mode: execute + approvedIds (the run-folder ids the human approved in chat). For each approved candidate, drive_trash_file({ fileId: <id> }). Trash nothing outside approvedIds. Report success/failure per-item, but never abort the batch on a single-item failure — log and continue. Then (sanity check) for one affected opp, drive_list_folder on its runs/ again and confirm the trashed names are gone.

Failure modes

• runs/ folder doesn't exist on a given opp — skip silently. Opps that never ran a /ace:run have no runs/.
• keep >= run count for an opp — that opp has zero candidates; omit it from the report.
• drive_trash_file permission error on a Shared-Drive folder — log as "trash failed — needs admin"; continue.
• keep == 0 — refuse. Retention prune always keeps at least 1 (the freshest run). Surface a clear error to the orchestrator; do not silently fall back.

Implementation notes for agents

• Use the Drive createdTime field — do not parse run-id slugs for timestamps. Run-id slug format has drifted historically; the metadata is authoritative.
• This skill does not touch opp.yaml, inputs/, eval-calibration/, or any other per-opp file. Only runs/<run-id>/ direct children get trashed.
• Run folders are deeply nested (per-phase verdicts, transcripts, screenshots). drive_trash_file on the run folder cascades naturally — no need to recurse.
• Drive trash is reversible for 30 days. If the human asks "what if I want it back," point them at https://drive.google.com/drive/trash and the run-id slug.

Related skills

• sweep-drive handles orphan opp folders (the level above this skill's targets). The two are complementary: opp-runs controls retention inside live opps; sweep-drive removes opps that no longer match the live-set.
• sweep-live-set is not an input here.