[jayleaton]

Summary

In-process agents saw two scheduling tool sets with different powers. The unprefixed set is now the one set; the t3_* set stays for outside clients.

 schedule_task          + advertises {type:'cron',expression,timezone}; "enabled=true … restores an archived one"; "For an agent, pass profileId"
                        + update now passes freshSession through (was dropped by the handler)
 list_scheduled_tasks   + includeArchived (default false)   result + archived: {reason, deletesAt}
 run_scheduled_task_now + scheduledTaskId (taskId kept as the older name)
 delete_scheduled_task    unchanged
 in-process t3_list/create/update/delete/run_scheduled_task
                        - full gateway descriptions
                        + "Deprecated: use `schedule_task` with `profileId`." (one line each, one release)
 gateway t3_create_scheduled_task
                        + idempotencyKey → stable commandId, so a retry returns the same task
 timeline labels        - update_scheduled_task (retired)   + the five t3_* schedule tools

Docs: remote-access.md no longer says linked machines can manage scheduled tasks (they can't, per internals/remote.md and linkOrigin.ts), and says in-T3 agents use schedule_task + profileId; orchestrator-mcp-server.md covers cron, agents, archive/restore and the deprecated aliases. agentContextBudget.json: tool list 60396 → 60391 bytes.

Sits on #222.

Evidence

Unit tests, written first. With the implementation reverted to #222 (patch kept aside, then reapplied): 6 failing —
adds every tool … lifecycle alias (deprecated descriptions), lists archived tasks only when asked, a caller cannot rewrite … (freshSession dropped), run_scheduled_task_now takes scheduledTaskId, labels the T3 Agents schedule tools, gateway turns runAt or cron into a schedule (idempotencyKey). After: server MCP 103/103, shared 5/5, mcp-gateway 115/115. vp run check:changed: all 10 steps pass.

Evals (evals/mcp-tool-use/cases/pull-requests.ndjson, three new bad cases):

Case	Must	Fails on old text because
schedule-cron-month-days	schedule_task with schedule.type: cron, expression ~1,15, never fixed_time	old schedule description lists only interval / fixed_time / once / webhook
schedule-agent-uses-schedule-task (agentsTools: true)	schedule_task with profileId: reviewer, never t3_create_scheduled_task	old t3_create_scheduled_task description is the only one that mentions agents + cron
schedule-restore-archived	list_scheduled_tasks {includeArchived: true} then schedule_task {scheduledTaskId, enabled: true}, never a new task	includeArchived does not exist on the old text

• --dry-run builds all three on the old text and the new text (exit 0). mcp-tool-use/run.test.ts: 481 passed.
• Paid runs are out of scope for this PR, by maintainer decision. The three cases stay in the file for the next eval run.

Merge Danger

Door: two-way. Tool text and an optional parameter; no stored state changes. An outside agent that calls an in-process t3_*_scheduled_task tool still works, it only reads a shorter description.

Blast Radius: agents

Every agent's tool list changes, but stays under main's size (60391 bytes against 60396). Agents that relied on list_scheduled_tasks returning archived tasks now need includeArchived. The three eval cases above cover the new text.

Implemented by Claude Opus 5.5 in Claude Code (T3 Code).

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	21.5 KiB	—	24.4 KiB	✅
Codex	Live turn messages	—	1	—	8	✅
Claude	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	21.9 KiB	—	24.4 KiB	✅
Claude	Live turn messages	—	1	—	8	✅

Baseline: unavailable · PR result: e3fbc4e · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[jayleaton]

VERDICT #224: CHANGES

No new handler/schema regression found: both run ID names remain accepted and freshSession reaches the update service. Resolve the required red budget check; an exact smaller wording is inline. Reviewed head a8bd85b.

Repository check:changed against #222 passed, including 37-package typecheck and 11 test files. Scheduling eval dry runs built all 3 added cases; eval grader tests passed (481). Paid behavior evals were not run, as the PR body explicitly scopes those out.

[jayleaton]

Fixed in e7d0f17: rebased on #222, and the scheduling tool text is trimmed to 60391 bytes, under main's 60396.

[jayleaton]

VERDICT #224: MERGE

Reviewed e7d0f17 against e1e6d06. check:changed passes: formatting, lint, both knip passes, GraphQL, 37-package typecheck and 11 focused test files. The budget comparison passes without a label override: 60391 bytes against main’s 60396.

The stdio owner-handover test passed 5/5 isolated runs on this head and 3/3 on its base. The failing CI log reports “Launcher exited: ... read ECONNRESET”. The test, transport, launcher, owner, bridge, entry point, server registration loop and dependencies are unchanged; this test lists tools and replaces an owner, never calling scheduling. The scheduling catalog gains an optional create parameter/description, with no handover protocol change. I found no functional link to this diff; the unchanged transport treats ECONNRESET as fatal, consistent with a pre-existing handover flake. The CI failure did not reproduce locally.

This layer passes review independently. The stack remains blocked by the new cleanup-reservation finding on #222.