Fix missing DSH schedule tools: presets and subagents

When the model will not create a reminder, the first question is not “can the model do it?” but “does this agent have the schedule tools?” DeepSeek Harness declares the four reminder tools per preset: standard, cordis, and ptc have them; minimal does not; delegated subagents are denied. Confirm the tool is present before debugging the prompt.

DeepSeek Harness: Check the preset first

Different presets in one DSH instance expose different capabilities. Compare:(source)

With minimal, the model cannot see schedule_create and has no reason to know reminders exist. Switch to a preset with the tools or start a new session using standard.

After switching, do not immediately ask for a reminder; ask the model to list available tools and confirm all four names.

DeepSeek Harness: Check whether a subagent is calling

A subagent may look able to call the tool and still be rejected at the tool layer. The three presets’ delegation rows deny the four tools. Symptoms:(source)

1. Main agent can create — ask in the main session. Expected: the task appears on the automation page.
2. Subagent cannot create — delegate the reminder task. Expected: the subagent reports no permission or no tool.
3. A constructed call is still rejected — the tool layer denies it. Expected: no silent success.

The right split: the subagent calculates timing or drafts the instruction; the main agent calls schedule_create.

DeepSeek Harness: Check the provider and time arguments

Tools present and main agent active still leaves two common causes: no model provider or invalid time arguments. Check:(source)

1. Provider configured — the official docs require a model provider before asking for reminders. Expected: normal chat works.
2. One-shot delay is positive integer seconds — do not pass “half an hour” as a string to the tool. Expected: a valid number.
3. Fixed interval is at least one minute — shorter intervals are invalid. Expected: use the smallest allowed unit.
4. Daily and weekly include an IANA zone — for example Asia/Shanghai. Expected: do not rely on a browser guess.
5. Cron has five fields and a zone — not six fields, not a missing zone. Expected: normalized storage.

If the model only says “creation failed” without the tool output, ask it to paste the original error.

DeepSeek Harness: Three creation paths

Chat, page creation, and update all land on the same task data, but they fail in different layers. Compare:(source)

1. Main-session chat — the model calls schedule_create. Expected: most direct, but depends on preset and provider.
2. Automation page “New” — opens a session with the request prefilled. Expected: good for manually specifying a frequency.
3. schedule_update on an existing task — keeps the ID and delivery records. Expected: good for adjusting a known reminder.

If chat fails but the page succeeds, the problem is the agent preset or tool declaration, not the Schedule service.

DeepSeek Harness: Reverse-check from the page

A working page does not mean a working agent; separate “service” from “agent”. Steps:(source)

1. Create a reminder on the page — complete the flow from “New”. Expected: the task saves, proving the service works.
2. List tools in the target session — ask the main agent for the four names. Expected: the missing name identifies the preset problem.
3. Check the preset name — standard, cordis, ptc, or minimal. Expected: match the matrix.
4. Switch presets and retry — start a standard session and create again. Expected: success confirms the original preset lacked the tool.

If the page also cannot create, the service layer is the problem. If the page works but chat does not, inspect the agent preset and delegation.

DeepSeek Harness: Three requests to try verbatim

Use the simplest requests first to rule out phrasing. Three:(source)

1. One-shot — “Remind me to check the build in 3 minutes.” Expected: schedule_create runs and the task appears.
2. Daily — “Remind me every day at 23:00 Asia/Shanghai to check the weather.” Expected: a daily rule with an explicit zone, not a one-shot.
3. List — “List all current reminders.” Expected: schedule_list returns active tasks only.

If the first works and the second fails, the problem is time-zone or frequency arguments. If all three fail, return to the preset and tool declaration.

A verification sequence

Check tool presence → caller → arguments → delivery; do not skip steps. Sequence:

1. List tools — confirm the four schedule names in the main agent. Expected: all present.
2. Create a one-shot — ask for a reminder a few minutes out. Expected: the page shows the task.
3. Check state — the task is active, not already not-running. Expected: it can be scheduled.
4. Wait for delivery — the original conversation inbox receives a follow-up. Expected: the record proves inbox persistence, not model completion.

If step 1 fails, return to the preset. If step 2 fails, inspect arguments. If step 4 fails, check whether the host is running; scheduling stops with the host.

DeepSeek Harness FAQ

Does an accessible page mean the tools are present? No. The page only proves the Web profile includes Schedule. Tool exposure depends on the agent preset.

Can I add the tools to the minimal preset? You can change the preset declaration, but that is a configuration-layer change. Switching to standard or cordis is simpler.

Is “no permission” from a subagent a bug? No. The delegation rows deny the four tools by design. Let the main agent create the reminder.

The short version: check “is the tool present”, then “who is calling”, then “are the arguments valid”; reversing the order makes you debug the wrong layer.