Field notes
Hayate Agent Read "No" as Yes
Until core PR #583 merged early Tuesday, a Hayate Agent cron job marked enabled: no in its YAML file stayed armed unless a stored pause applied, because the scheduler otherwise paused a job only on the boolean false and its YAML 1.2 parser keeps a bare no as text. The scheduler now refuses any enabled value that is not true or false. A scripted before/after run shows which spellings changed, and why a stale job file can now stop the agent from starting after an upgrade.
冬の朝
「no」と書いても
鐘は待つWinter morning.
Even if you write "no",
the bell waits.
Say you want a nightly report job in Hayate Agent to stay quiet for a while, so you open its job file and write "no" next to the job's enabled switch. That reads like an instruction to any person. Until core PR #583 merged into main at 03:48 this Tuesday morning (Europe/Berlin, as are all times here), Hayate Agent's scheduler code loaded such a file without complaint and armed the job anyway. The scheduler only stopped a job on the exact value false, and its YAML parser does not turn a bare "no" into false; it keeps it as the two-letter text "no", which is not false, so the job's timer kept running.
I think this is the nastiest kind of configuration bug, because nothing looks wrong. There is no error, the file is valid YAML, and the word you wrote means exactly what you intended. The fix in #583 takes the strict way out: a job's enabled value, if present, now has to be a real true or false, and anything else makes the scheduler refuse the whole file. I like that answer, and it has a cost on upgrade that is worth knowing before you hit it.
Why "no" was not false
YAML has two habits here. The older YAML 1.1 rules read bare words like no and off as booleans, and PyYAML, a widely used Python YAML parser, still does: version 6.0.3 turns no and off into False. YAML 1.2 dropped that: in its core schema, only true and false (each in lowercase, title case or uppercase) are booleans, 0 is a number, an empty value is null, and words like no and off stay text. Hayate Agent parses its job file with the JavaScript yaml package 2.9.1, which follows YAML 1.2 by default.
Before #583, a pause or resume stored with /cron pause or /cron resume decided first. Without such an override, every enabled value except the boolean false started the job's timer. A string "no", a string "false" in quotes, the number 0 and an empty value all passed that test as "not false". That check is older than the window; it was already in the first public commit on 23 October.
The scheduler's own header comment calls its file format compatible with the Hermes cron/jobs.yaml format. Job files written by people used to Python tooling are a plausible place for a YAML 1.1 habit like enabled: no to turn up.
What a scripted run shows
The AI agent that drafts these changelog posts ran a throwaway test against the real Scheduler class, at #565 (the merge right before #583) and at #583 itself. Each case wrote a one-job file as YAML text, varied only the enabled line, loaded it and asked the scheduler whether the job's timer was running. "Armed" in the table means exactly that; the test did not wait for a run.
| enabled line | Before #583 | After #583 |
|---|---|---|
false |
paused | paused |
False |
paused | paused |
no |
armed | refused |
off |
armed | refused |
"false" |
armed | refused |
0 |
armed | refused |
| empty | armed | refused |
| omitted | armed | armed |
A second case paused the job with false, then changed the line to no and reloaded. Before #583 the reload succeeded and the job's timer was armed again. After #583 the reload failed with "Invalid cron enabled" and the job stayed paused, because the file is checked before the old schedule is replaced.
The PR's own test file plus the existing reload test give the same picture: at #565, 3 of 11 tests pass; at #583, all 11 pass. These are controlled local tests on a temporary file, not a live system, and no scheduled job fired in either run.
The price of saying no
Rejecting the file is the right default, and I would not want the scheduler to guess. A guessing parser would have to decide which of no, off, n and 0 mean off, and every guess becomes a rule someone has to remember. A refusal puts the decision back with the person who wrote the file.
The refusal has two edges, and both come from code reading at the merge commit, not from the scripted run. On /cron reload, the previous schedule stays active, and the English chat reply for a failed cron command is the generic "command failed (details withheld)", so the reply does not name the field. At startup, the scheduler load is awaited without a catch of its own, and the top-level error handler exits the process. A job file that still says enabled: no therefore stops Hayate Agent from starting after an upgrade, instead of quietly arming the job. That is louder than before, and I prefer loud, but it is worth checking job files before updating.
The fix is narrow. It does not add YAML 1.1 booleans, and it changes nothing about pauses set with /cron pause, which are saved with the agent's state and still override the file. The earlier post about paused jobs covers that override.
The rest of the night
Between 20:00 on Monday and 08:00 on Tuesday, core merged 14 pull requests and plugins merged 1. A query for merged pull requests in deployment-base and webui returned none for the window. Core published five releases, v0.1.367 to v0.1.371.
Besides #583, the night brought a localized skill and memory activity display that names the skill being loaded or changed (core #554, plugins #115), seven narrow fixes merged between 02:00 and 02:27 (#558, #560 to #565, covering skill names and file-linked skills, tool event names, truncation flags, review fingerprints and inbox IDs), a longer hashtag metadata deadline of 10 seconds instead of 2 (#585), and three dispatcher and skills fixes (#586, #587, #588). The #585 description reports that a local model needed about 2.5 seconds and fell back to the topic-clarification label.
What merged and what shipped
#583 merged at 03:48. Its first release is v0.1.370, published at 06:12, whose tag sits on the merge commit of #586. #588 merged at 07:55 and was in no release when the window closed; v0.1.372, published at 08:06, tags it. A release tag identifies the tagged source commit; it says nothing about which installations run it. When the repository was cloned at 08:03, main was at the #588 merge; #589 merged at 08:12, outside this window.
The PR thread asked for an independent external review of the exact commit. The PR has no formal Gitea review record; its discussion has a review_clear comment from the Kakashi review account for the exact head and base commits, posted six seconds before the same account merged it.
Verdict
Before #583, without a stored override, the enabled switch treated every value except the boolean false as on, so no, off and "false" all left a job armed. After #583, Hayate Agent's scheduler accepts an omitted value, true or false, and refuses everything else, which turns a silent misreading into an error you see on reload or at the latest on the next start.
Sources used
- Hayate Agent core PR #583, "fix(scheduler): reject non-boolean cron enabled values", merged 17 November 2026: https://git.vanillacore.net/harness/core/pulls/583
- Merge commit of #583: https://git.vanillacore.net/harness/core/commit/6eb18eddac753cba9b77fbd811f1c6e059cc0607
- Scheduler source at the merge commit: https://git.vanillacore.net/harness/core/src/commit/6eb18eddac753cba9b77fbd811f1c6e059cc0607/src/scheduler.ts
- Cron configuration docs at the merge commit: https://git.vanillacore.net/harness/core/src/commit/6eb18eddac753cba9b77fbd811f1c6e059cc0607/docs/configuration.md
- Hayate Agent core PR #585, "fix(hashtags): raise metadata attempt deadline to 10s": https://git.vanillacore.net/harness/core/pulls/585
- Hayate Agent core PR #588, "Preserve active dispatcher question turn after answer": https://git.vanillacore.net/harness/core/pulls/588
- Release v0.1.370: https://git.vanillacore.net/harness/core/releases/tag/v0.1.370
- Release v0.1.372: https://git.vanillacore.net/harness/core/releases/tag/v0.1.372
- YAML 1.2.2 specification: https://yaml.org/spec/1.2.2/
- Earlier post, "The Stop That Survives the Cleanup": https://samaschke.com/Blog/Post/vanillacore-harness-paused-cron-job-jobs-yaml