Delete a single turn — by button and by voice ("scrap that") — because a noise message keeps shaping every later reply

[serge-ivo]

The ask

I want to be able to delete my last message, and any last message — in case I say something wrong
or it picked up noise and did something based on that. And a voice command like "scrap last
message" so I can do it by voice instead of deleting manually.

Today

• DELETE /v1/instances/:id/messages is clear-the-whole-thread (the console's Clear chat,
  InstanceDetail.tsx:800). There is no per-message delete anywhere.
• The voice vocabulary is repeat | mute | unmute | exit | next — no undo.

So the only way to remove one bad turn is to destroy the entire conversation.

Filed as one ticket, two triggers

The button and the voice command are the same operation invoked two ways — the same reasoning that
made #277 ("next") and #279 (transfer) share one switch primitive. Building them separately would
produce two delete paths that disagree about what a "turn" is. Happy to split if preferred.

Delete the TURN, not the message

A user message never stands alone: the assistant's reply and the tool-log system message are
written in the same turn and are meaningless without it. Deleting one line would leave an answer to
a question nobody asked.

Unit of deletion = the exchange: the user message, its assistant reply, and the tool-call system
message from that turn.

The part that must not be got wrong

Deleting the message does not undo what it caused.

This is the whole motivation — "it picked up some noise and did something based on that" — and the
naive implementation makes it worse. If noise triggered start_work, removing the transcript
leaves the run going, with nothing left on screen explaining why the agent is working. That is #251
inverted: there the side effects persisted while the record vanished by accident; here we would do
it deliberately.

So deleting a turn must:

• Say what it caused. The tool calls are already recorded in the turn's system message and in
  agent_events; the confirmation should name them ("this turn started run 4a03… and wrote 2
  memory entries").
• Offer to cancel what is still running. A live loop run has requestCancel already.
• Not silently orphan durable effects — a write_memory, a created ticket, a filed GitHub
  issue. Deleting the message cannot retract those, and pretending otherwise is the failure mode.

If nothing can be undone, say so plainly and let the user delete anyway. The goal is an honest
record, not a fake one.

Soft delete, with a window

The undo needs its own undo. The voice trigger especially: "scrap that" arrives as a voice turn,
so the same noise that caused the problem can plausibly produce the command — and #332 is the
standing evidence that silence transcribes into real-looking phrases.

Mark deleted, hide from the thread and from the model's context, purge after a short window (and
purge the audioKey blob with it, as Clear chat already does). A "Deleted — undo" affordance for a
few seconds costs nothing and makes both directions recoverable.

The voice command

Add scrap to VoiceCommand with the usual treatment: a SCRAP_BY_LANG phrase table, custom-word
override, and the existing whole-utterance or trailing-phrase matching so "we should scrap that
approach" stays a message.

Phrases worth including: "scrap that", "scrap last message", "disregard that", "delete that",
"ignore that".

Note #334: normalizeTranscript does not strip hyphens, so "scrap-that" would fail to match. Fix
that first or this command inherits the same intermittency.

Why it matters beyond tidiness

The thread is the context. A phantom turn is not just visual clutter — it is fed back to the
model on every subsequent turn, so a noise message keeps influencing answers until the conversation
rolls over. Deleting it is the only way to stop that.

Verification

• Deleting a turn removes the user message, the reply and the tool log together.
• The confirmation names any durable effects, and offers to cancel a live run.
• A deleted turn is absent from the next request's context.
• "Scrap that" as a whole utterance deletes; "scrap that approach" does not.
• Undo restores the turn intact, including its audio, within the window.

[serge-ivo]

Shipped in c158425 (the button) and a617706 (the voice command). CI green on both.

What a delete removes

The exchange, not the line: the user message, the agent's reply and the tool-log system
message written with it. Pointing at an assistant or tool message resolves back to the ask it
answered; the span ends at the next user message. One pure rule in
workers/api/src/lib/chat-turns.ts, not a walk open-coded twice.

Three things go together, exactly as Clear chat already ties them: the message record, its
audioKey blob in R2, and the dictation stored beside it (#319). The dictation needs no
handling at all — it lives ON the message, which is why it was put there rather than in a second
store with a second retention rule. The audio ids are returned for the route to delete, never a
prefix sweep (the Coder Co-pilot's recordings share that prefix).

Not removed, deliberately: conversation summaries and the facts extracted from them. A summary is
an aggregate over twenty messages and cannot be un-mixed, so the console says so rather than
implying the turn is unmade.

The phrase rule, and what it must not fire on

scrap is matched whole-utterance only, regardless of word count — a new { whole: true }
mode on phraseMatchesTranscript. The usual multi-word rule (match as a whole-word run anywhere)
does not survive this word list:

• "don't scrap that, keep it" — the user refusing the action
• "scrap that idea and let's move on"
• "we should scrap that approach"
• "ignore that warning for now"

All four stay messages; each is pinned by a test. Bare "delete that" is not in the table at all
— a Coder subscriber says it about a file, and no matching rule can separate the two readings.

Same reason, second narrowing: it is judged only on FINAL utterances. Three of the five matcher
call sites in use-voice.ts see interim transcripts (the gate scan, the partial path, and the
always-on control listener, which checks interim as well as final), and a partial of "scrap that
idea and let's move on" is momentarily exactly "scrap that". Those three withhold the
canScrap flag — the same canSwitch-style gate next uses. Cost: "scrap that" said while the
agent is talking does nothing until the mic reopens. Missing a delete is the cheap failure.

splitTrailingCommand never learns the word either: a whole-utterance command has no message half
to keep, and "let's rewrite the parser, scrap that" must not both delete and send a truncation.

Confirmation, not an undo window

Both triggers confirm before, and there is no soft-delete window. That is a deliberate
departure from the ticket's "soft delete with a window", for one reason: an undo toast only helps
someone who is looking. The hands-free user this command exists for has their eyes elsewhere, so
a thing that expires in ten seconds protects nobody. The gate belongs before — the same call #328
made for delete_supervision, and for the same shape of risk (nothing errors afterwards, so a
mis-fired delete looks exactly like an intended one).

Three narrowings carry the safety:

1. The voice confirmation QUOTES the turn. The command arrives through the same unreliable
   channel that produced the problem it is fixing; [bug] Silence is transcribed as the agent's own name — proper nouns in the Whisper bias list come back verbatim #332 is standing evidence that silence
   transcribes into real-looking phrases. Showing the words about to be destroyed is what turns a
   yes/no into a decision.
2. Voice targets the LAST turn only. There is no reliable way to say "the one before that" out
   loud. The button reaches any turn, because the user pointed at it.
3. Both confirmations name what the turn DID and state plainly that deleting the transcript is
   not a retraction — anything it started keeps running, anything it wrote stays written. When a
   loop is live, the button offers to stop it. If nothing can be undone, it says so and lets you
   delete anyway.

#334's normaliser already strips hyphens, so "scrap-that" matches — pinned by a test, since a
destructive command cannot afford the #334 intermittency.

Also

Ten languages, scrapWords custom override in Settings → Voice beside the other four, and
DELETE /v1/instances/{instanceId}/messages/{messageId} documented in store/openapi.yaml.