The job had just been renamed. Listing the notepad under the new name said it was empty. The prompt string the scheduler builds for that same job still opened its scratchpad with cursor 17, and listing the old name still returned cursor 41. Nothing in hermes cron doctor mentioned either row. The notepad command had accepted every string it was given, and the run path reads only one of them.
That split is the contract in Hermes Agent v0.22.5+471.g5894e75, local commit 5894e750c3db, measured on 15 November 2026 on Linux x86_64. The CLI help calls the first argument a job ID. Pause, edit, and remove also accept a name, and they resolve it. The notepad subcommand does not. The official cron guide, fetched the same morning, does not mention the notepad at all, so the help line and the other subcommands are what an operator has to go on.
One paused job, three stored strings
The directory under test was a new Hermes home. The operator home was not that directory. No gateway was installed for it, and the job was created paused so a tick could not fire it. Delivery was local. The schedule was a cron expression that was never resumed. The create line was:
hermes cron create --paused --name watermark --deliver local "0 0 1 1 *" "Record the watermark and stop."
The reply named the id bd0dd75334ff and the name watermark, and it said the job was created paused. The stored record had no_agent false and an empty skill list. The id is twelve hexadecimal characters. In this commit, create draws that value from a random UUID and keeps it when the name changes. The name is a label. The two are not interchangeable keys, even when you only ever type one of them.
Three writes followed, each with the same key and a different cursor:
hermes cron notepad watermark set cursor 41
hermes cron notepad WATERMARK set cursor 99
hermes cron notepad bd0dd75334ff set cursor 17
Each exited 0 and printed that the key had been set for the string on the command line. Listing each string back returned only its own value. The lowercase name held 41, stamped 2026-11-15T11:07:08.291356-05:00. The uppercase name held 99, stamped one second later. The id held 17, stamped 2026-11-15T11:07:10.038786-05:00. Three rows, one key name, three job-id columns. A later list of one string does not show the other two.
That is the whole storage rule. The set path validates that the job-id argument is non-empty, then upserts (job_id, key). It does not open the jobs file. A string that matches no job is still a successful write. A string that matches a name is still stored as that string, not rewritten to the id the name belongs to.
The commands beside it do resolve a name
The contrast is the rest of the cron CLI in the same binary. hermes cron pause WATERMARK exited 0 and printed Paused job: watermark (WATERMARK). Name match there is case-insensitive. Exact id wins if you pass an id. An ambiguous name raises instead of picking a row. None of that lookup runs for notepad.
hermes cron edit watermark --name tide-mark updated job bd0dd75334ff and left the schedule alone. After that edit, a list under tide-mark was empty. A list under watermark still returned cursor 41, with the original timestamp. A list under the id still returned cursor 17. Renaming does not move notepad rows, and it does not warn that rows exist under the old name. The new name starts empty even when the job you just renamed had a scratchpad an operator would recognize by the old label.
hermes cron doctor then printed that it found no issues, and that no active jobs were configured. The job was paused, and this doctor only walks active jobs. It also never opens the notepad database. Silence here does not mean the three rows agree. It means this check does not look. This run did not enable the job and ask doctor again, so it does not say what doctor prints for an overdue active job. It does say that a name-keyed notepad is outside the checks that command currently performs.
What the scheduler would actually prepend
The agent path builds the job prompt in _build_job_prompt. That function asks the notepad renderer for job["id"] and, when the result is non-empty, puts it in front of the job prompt. The scheduler calls that function when it is about to run an agent job. This measurement called the same function on the stored record after the rename and after one more write. It did not start a gateway, did not tick the scheduler, and did not call a model. The object under test is the prompt string, not a reply.
The extra write put a two-line value on the id. The first line was the literal line-one. The second line was a Markdown heading that said injected heading. The command was a notepad set of the key note against the job id, with that newline inside the value.
get printed both lines. The assembled prompt contained cursor: 17. It did not contain cursor: 41 or cursor: 99. It did contain the two-line value, and the second line was ## injected heading, sitting in the prompt ahead of Record the watermark and stop. The renderer joins each stored value onto a dash line. It does not quote, escape, or reject a newline. A value with a newline is not a field in a list. It is more prompt text. In the list command, the same newline makes the updated: stamp look like it belongs to the second line of the value, because the printer does not bracket the value either.
The section also tells the model how to update the scratchpad, and the example it prints uses the id, not the name. A cron agent that copies that example writes a row the next run will see. An operator who copies the name from the job list, because pause and edit accepted that name a moment earlier, writes a row the next run will not see. Both writes succeed. The difference shows up only when you render the id, or when you compare the three lists.
A no-agent job does not take this path. The script short-circuit never calls the prompt builder, so a notepad on a script-only job is not injected into a model prompt by this function. This job was not a no-agent job. A script-only notepad was not fired here. Do not treat the prompt excerpt above as evidence about script stdout.
Removal clears the id and leaves the name
hermes cron remove tide-mark resolved the new name, removed job bd0dd75334ff, and cleared notepad rows for that id. The id list was then empty. The lowercase name still listed cursor 41. The uppercase name still listed cursor 99. The jobs file had no remaining job. The scratchpad file still had two rows whose job-id column had never been an id.
That is the cleanup rule in this commit. Remove resolves a name or an id to the canonical id, deletes the job, and deletes notepad rows for that id only. It does not search the notepad for the name it was given, the previous name, or any other case. If the clear itself throws, the removal still returns success and the failure is a debug log. This run did not hit that failure. The id row was gone. The name rows were not.
A second, empty Hermes home, asked to list the same id, also printed that the notepad was empty. That is the same sentence the new name produced before the remove. An empty list means no row for that string in this home. It does not mean the job exists, and it does not mean some other string for the same job is empty. The other home also did not see the name rows. The database is profile-local. A gateway or a shell with a different HERMES_HOME is a different notepad file, which is what you want, and it is also why a check run in the wrong home will tell you the scratchpad is empty when the job’s home is not.
The size cap is a different refusal
The command will refuse a value. A 16,385-byte value of the letter x, written against the id before the remove, exited 1 with value too large (max 16384 bytes per key). The cursor row’s timestamp did not change, and no big key appeared. The cap is on the UTF-8 size of one value. The source also caps a key at 128 characters and a job’s keys plus values at 64 KiB, and an over-cap write is supposed to leave the store untouched. This run hit the per-value cap and saw the store unchanged. It did not fill a job to 64 KiB, and it did not try a 129-character key. Those two limits are in the module. They are not a second measurement.
The cap does not make a name into an id. A short value under the job’s name is accepted. The refusal you can see in the exit code is about size, not about whether the next run will read the row.
Verify the id before the next run
Use a spare Hermes home, or a paused job you created for the check. Do not point these commands at a production id until you have printed the id from hermes cron list and decided which row you mean to change. Removing a job to see what remains is a lab step. On a job you still want, list first and delete only a leftover key.
Create or identify the job, and copy the id from the first column of hermes cron list, or from the create reply. The name on the next line is the label pause and edit will accept. It is not the notepad key the scheduler reads.
Set the cursor with that id. List the id and confirm the value and the timestamp. Then list the name you would have typed from habit, including a different case if you have used one. Those lists should be empty. If they are not, those rows are invisible to the next agent run and will survive a later remove. Delete them with the notepad delete action on that exact string, or leave them and know they are orphans.
Rename only after that check. List the id again. The value should be unchanged, because the id did not change. List the old name and the new name. The old name still has whatever you wrote there before. The new name is empty unless you wrote that string too. The prompt builder does not follow the name.
If a value must survive into the next prompt as data, keep it on one line. A newline is concatenated into the prompt. A line that looks like a heading is a heading in that prompt, ahead of the job text. Do not store a page, a ticket, or a monitor diff in the notepad if the next run should treat the scratchpad as a cursor rather than as instructions. The write path is the CLI. A cron agent with a terminal can run that CLI. Whatever it writes under the id is the prefix of the next agent run.
After you remove a job, list the id and every name you used. The id list should be empty because remove cleared it. A name list that still prints a value is a row remove did not own. This home’s file is the one that matters. A list from another home only tells you about that other file.
This run did not resume the job, did not deliver a report, and did not ask a model what it would do with the injected heading. The measured boundary is earlier than a reply: which string is stored, which string the prompt builder reads, which string remove deletes, and which refusal the size cap actually prints.
Sources
- Hermes Agent v0.22.5+471.g5894e75, local commit 5894e750c3db, Linux x86_64, 15 November 2026, US Eastern. Isolated Hermes home.
hermes cron create --paused --name watermarkcreatedbd0dd75334ff. Notepad writes underwatermark,WATERMARK, and the id stored three differentcursorvalues.hermes cron pause WATERMARKresolved the job.hermes cron edit watermark --name tide-markdid not move rows._build_job_prompton the stored record includedcursor: 17and a newline value whose second line was a heading, and did not include the name rows. A 16,385-byte value exited 1 with the 16,384-byte cap and left the existing keys in place.hermes cron remove tide-markcleared the id and left both name rows. A second home listed the id as empty.hermes cron doctorreported no issues and did not read the notepad. - Installed source at that commit:
cron/notepad.py(set_note,render_notepad_section, the 16 KiB and 64 KiB caps),hermes_cli/cron.py(cron_notepaduses the argument string;cron_doctordoes not query the notepad),cron/jobs.py(resolve_job_ref,remove_jobclearing the canonical id, job ids from a twelve-character UUID slice),cron/scheduler_prompt.py(_build_job_promptrendersjob["id"]), andcron/scheduler.pycalling that builder for an agent run. - Scheduled Tasks (Cron), fetched 15 November 2026. The page describes create, pause, edit, remove, no-agent mode, and delivery. It does not mention the notepad. The CLI help text for the notepad subcommand is the line that calls the argument a job ID.