[FEATURE] Auto-archive (auto-close tab) for scheduled task / routine run sessions after completion

I also have this issue. I have avoided running as many routines as my colleague - his computer started failing and I would not be surprised if it was due to the active running sessions from all routines he's got setup.

The routine feature is great - I really love the concept - I'm just feeling it is a unpolished diamond - and this type of fix would polish it alot for me.

Adding a scale data point and one mechanism detail I don't think is in this thread yet.

Scale. I run 93 scheduled tasks on the desktop app (macOS). A list_sessions pull this week returned 453 sessions, 421 of them completed routine runs — 404 of those idle more than a day, the oldest 65 days. Nothing was deleted or misconfigured; that is what about two months of a normal routine fleet accumulates. The 18-runs/day case in the original report is the small version of this.

It also breaks the assistant's own session tooling. At that many live sessions, mcp__ccd_session_mgmt__list_sessions returns roughly 88k characters at limit: 250 and 160k at limit: 1000, both past the tool-result token cap, so the result is written out to a file instead of returned in context. The documented triage workflow (list → rank → archive) can't run in context at all once a routine fleet has been going for a couple of months.

New mechanism detail — the user-level allowlist doesn't reach it either. This issue already establishes that a task-level approvedPermissions entry for archive_session still prompts. The same holds one layer up. ~/.claude/settings.json here contains:

"permissions": { "allow": ["mcp__ccd_session_mgmt__*"], "defaultMode": "auto" }

and every archive_session call still prompts. The dialog says so itself — "This tool requires explicit approval regardless of permission mode" — and offers only Deny and Allow once, with no "always" option. So neither the task-approval layer nor the user settings allowlist touches it.

Consequence for a backlog that already exists. With no bulk form and no allowlist path, clearing an accumulated fleet costs one human click per session. I archived 51 by hand before stopping, with 353 left. An auto-archive setting fixes the future and leaves the existing ~400 untouched.

So there are two asks here:

1. The retention / auto-archive setting proposed above. Every routine fleet accumulates this way, so I'd argue it should ship on by default.
2. A way to clear a backlog that already exists: either let archive_session be pre-authorized like other MCP tools, or add a bulk archive_sessions(session_ids[]) that takes one approval for the batch.

Happy to supply the raw counts if they're useful.

Cloud Routines: no way to prune run history, and no viable workaround

Summary

A recurring Routine accumulates run history indefinitely. There is no bulk
delete in the UI and no API path at all. On a routine that has fired daily
since April, the only option is clicking away several hundred runs one at a
time. In practice only the latest one or two runs are ever useful for me.

What's missing

No way — UI or API — to delete a Routine's old runs. Concretely, from an
agent session:

• list_triggers / delete_trigger operate on routines, not runs.
• A routine record exposes only last_run, a single entry. The rest of the

history isn't enumerable.

• Routine runs are separate sessions (cse_...). list_sessions excludes

trigger-fired runs by default, and its tags filter returns
tags filter is not currently available for in-session callers — so the
runs can't even be listed, let alone deleted.

• There is no delete-run / purge-history tool.

Requests, roughly in order of usefulness

1. Retention setting per routine — "keep last N runs" (N=1,2,5,10) with

older runs pruned automatically. This alone solves it.

1. Bulk-delete in the UI — select-all / "delete runs older than <date>".
2. API parity for run history — a way to list and delete a routine's runs.
3. Make create_trigger config-complete — sources, model, outcome

branches, allowed_tools. Independently useful, and it would at least make
the clone-and-delete workaround real.

Environment

Claude Code on the web (web_claude_ai origin), container CC 2.1.259,
anthropic_cloud environment. Routines created in the web UI, fired on a
0 3 * * * cron since 2026-04-27.