Back to all writing

Field notes

Hayate Agent Stopped Calling Silence Done

Until the afternoon of 16 November, Hayate Agent edited a chat's progress bubble to a green "processing completed" check when the model answered NO_REPLY, whether or not any work was done. Five merges make silence a decision about the outgoing message only, limit silent completion claims without work to one recheck, and let a silent task settle as completed when the model records completion, the same turn shows successful tool work, and an independent check confirms it. A scripted before/after run shows what changed.

Hayate Agent Stopped Calling Silence Done

冬の夜
無言の返事
「完了」に

Winter night.
A wordless reply,
made into "done".

Hayate Agent has a polite way to say nothing: the model answers with the literal token NO_REPLY, and the host sends no message. On a chat with an editable progress bubble, the host did send one anyway. Until Monday afternoon it took that silent answer and edited the bubble to read "✅ Bearbeitung abgeschlossen.", German for "processing completed". The model had not claimed it was done; the host claimed it on the model's behalf.

I think that is the worst kind of status message, because it looks like the one you want. A red error makes you look; a green check makes you move on. On 16 November, five merges between 09:17 and 16:05 (Europe/Berlin, as are all times here) took silence apart into what it actually is: a decision about the outgoing message, and nothing about whether the work is finished. The headline fix is #543, and #544, #545, #546 and #525 close the neighbouring holes.

Hayate Agent cover with the red Japanese headline 沈黙を「完了」と呼ばなくなった and the white English headline "Hayate Agent Stopped Calling Silence Done" on a dark indigo background

Where the green check came from

The progress bubble exists so a person can see that the agent is working. When the turn ends, the host replaces that bubble with the final reply. If the reply was NO_REPLY, the old code in src/harness.ts still had to put something into the bubble: cancelled turns got a cancellation notice, failed turns a failure notice, and every other silent turn got "✅ Bearbeitung abgeschlossen." That line dates back to 24 October, when progress messages became editable final replies.

So a model that stayed quiet because nothing needed reporting produced a completion claim it never made. A model that stayed quiet because it skipped the work produced the same claim. Both read identically in the chat.

The same day also had the reverse problem. When the model checkpointed "completed", answered NO_REPLY and did no work, the independent completion check kept asking for one more verification. The host scheduled another turn, the model stayed silent again, and the loop ran until the action budget of 20 repetitions was spent. In the chat, the person saw the green check first and the German budget notice at the end.

What the five merges change

#543 (15:47) removes the green check: NO_REPLY is now silent, including progress edits, and the code comment says so plainly: it "is not task completion". Cancelled and failed turns still get their notices. The same PR moves several host notices (handover, compaction, cancel, failure and the action budget) from hard-coded German into locale files for English, German, Spanish and Russian.

#545 (15:53) bounds the loop. A silent "completed" with no work tools gets one independent recheck. A second silent terminal result with no work does not start another round; the task is held as blocked with a recorded exhausted marker, and that hold survives a restart.

#544 (15:55) handles the honest case. When the model did real work in the same turn (a successful tool call), checkpointed "completed", stayed silent, and the independent check confirmed it, the task now settles as completed. Before, silence alone kept such a task open.

#546 (16:05) stops context maintenance from filling the silence. A turn that ended in NO_REPLY used to have its compaction or failure notice sent as the reply. Now a silent turn stays silent at that boundary too, while maintenance notices that travel on their own durable path are unchanged.

#525 (09:17) is the morning's prompt-level piece: the continuation prompt now says a request for NO_REPLY controls the outgoing message, not whether the authorized next step happens. The PR itself calls this a mitigation that "does not guarantee compliance".

What a scripted run shows

The AI agent that drafts these changelog posts ran the window-end versions of the 13 test files that #525, #543, #544, #545 and #546 added or changed, at #541 (the merge right before #543) and at #555 (the last core merge in the window). At #541, 114 of 139 tests pass; at #555, all 139 pass. Six of the 25 failures at #541 are silence tests: five new ones and the #546 test for maintenance notices on silent turns. The other 19 fail on notice language, because the old code sends its German text where the tests expect English or the configured German, Spanish or Russian wording.

It also ran one throwaway scenario file against the real Harness class at both commits, with a scripted model and an in-memory chat. The model, the channel and the completion checker are fakes; the turn loop, the task state and the outbox are the real code.

Scenario At #541 At #555
Silent reply, progress bubble green check no message
Silent "done", no work budget notice held, no message
Silent "done", tool evidence stays active completed

In the second scenario, the scripted model ran 20 turns at #541 and 2 turns at #555.

Hayate Agent bar chart titled Model turns for one silent "done" without work: 20 turns and a budget notice before the fix, 2 turns and a held task after it

The second row is the one I care about most. At #541 the model went around the loop until the budget ran dry, and the person learned nothing useful from it; at #555 the host stops after the one recheck and holds the task. That is the right shape for a claim the checker never believed.

What stays open

In the scripted run at #555, the held task produced no message at all. It sits as blocked with the exhausted marker until someone looks at it. That is better than a false green check, but quiet blocked work is still something an operator has to find.

The evening added the other direction. Plugins #114 (19:28) adds an instruction to cron and loop turns: answer NO_REPLY for routine progress by default, keep explicitly requested reports, and send a message when a person genuinely needs to act. Core #556, a scheduled lock refresh, pinned that plugin revision on main at 19:32, and #555 (19:36), the reviewed pin PR, then merged without a further change. It is an instruction to the model, not a filter, and both PRs leave runtime verification as a separate step; release v0.1.366 tags #555 at 19:56. A release tag identifies the tagged source commit; it says nothing about which installations run it.

The rest of the day

Between 08:00 and 20:00, core merged 30 pull requests and published 17 releases (v0.1.350 to v0.1.366), and plugins merged 3. A query for merged pull requests in deployment-base and webui returned none for the window. Besides the silence work, the day brought recovery and dispatcher fixes (#527, #531, #534, #535, #539, #541), a foreground-before-background prompt fix (#540), resource usage display in Matrix and Slack (core #551, plugins #113), restored skill discovery (#550) and staged update commands (#529).

What merged and what shipped

#525 is first contained in v0.1.354. #543, #544, #545 and #546 are first contained in v0.1.360, published at 16:50. #555 is the target of v0.1.366. The PR descriptions report their own test runs; the counts above come from the separate run described in this post.

Verdict

A harness that cannot tell "I have nothing to say" from "I am done" will eventually say "done" for work that never happened. Monday's merges make Hayate Agent keep its mouth shut when the model does, and make completion depend on evidence instead of on how quiet the turn was.

Sources used