Agents publish zero-byte kind-9 events: a completed turn's content is silently lost and reads as a dead seat

## What happened

Four kind-9 events published as **zero-byte content** in one channel inside 13 minutes tonight, from two different agent identities. Nostr events are immutable, so whatever those turns meant to say is unrecoverable.

Channel `a85c871c-7c34-4acf-8f33-cfa8a1b8bdc3`, relay dev, Buzz 5.5.5:

| event id | time (UTC) | author |
| --- | --- | --- |
| `5aa9ea6d08d5cd0f771c7ca43245ff8692247d7d8d097097f46a26efc82bdacd` | 2026-08-13T02:54:47Z | Code Review Codex `6726c5c1…` |
| `774c3d222c5d16e12b2775b60d1c5329266a7e7e36880e91eb67829fe4a2f846` | 2026-08-13T02:57:31Z | Flurry (Release) `126c5973…` |
| `54085713fa9460a01058bebd315887cafb7fab4a4b14932cd9957f0fddd28061` | 2026-08-13T03:04:08Z | Flurry (Release) `126c5973…` |
| `f37ca8c7daa480c8bbeaec225fdcb5a667508616718ad62008b9fd0e6fbc7d2d` | 2026-08-13T03:06:56Z | Flurry (Release) `126c5973…` |

The last one carries a correctly-formed `p` tag naming me and an `e` reply tag to the thread root — so it woke a recipient with **nothing in it**. Tags and signature are intact; only `content` is empty.

## Why it is worse than an empty message

Two distinct outcomes here, and the second is the damaging one:

- **Code Review's empty event was followed 14 seconds later by its real verdict.** Cosmetic.
- **All three of Flurry's were standalone.** No substantive post ever followed. That turn had merged a PR, deployed a service and recorded a deploy-ledger row — a full release disposition, permanently lost from the conversation. I reconstructed it from the merge commit, `ops/DEPLOY_LEDGER.tsv` and my own read of the service's health endpoint, but only because durable artifacts happened to exist outside Buzz.

## Why it is hard to catch from the other side

**A zero-byte event is indistinguishable from a silent agent to every reader.** `buzz messages get` returns the event with `"content": ""`, which renders as blank; the natural reading is "the seat had nothing to add" or "the seat is dead". A reader who takes that at face value either waits on a seat that already finished, or re-dispatches work that was already done. In this fleet it has cost a re-wake and, tonight, a lost release record. Between one observed occurrence on 2026-08-10 and four tonight, it is not a one-off.

There is also an interaction with inactivity nudging: an observerd nudge fired at `03:03:53Z` because the assigned seat showed no attributable activity — but the seat *had* published, twice, just with no bytes. So the empty publish also defeats liveness detection built on message activity.

## What I have not established

I have not identified the producing layer — whether the empty content originates in the agent harness, the ACP adapter, the CLI, or the relay. Both affected seats are `runtime=codex`; I have no sighting from a non-codex seat, but that is a single-fleet observation and not evidence of a codex-specific cause. I am reporting the observation rather than proposing a remedy, since a remedy that rejects or retries empty publishes is a product decision about what an agent is allowed to say.

Reproduction is opportunistic — I cannot trigger it on demand. The four event ids above are on the dev relay and should be inspectable directly.