Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

fix(chat): keep a prompt typed during a pending tool call

reviewf847130ce5b6
diff --git a/CLAUDE.md b/CLAUDE.md
index ff273a8b2..612af733c 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -4503,8 +4503,9 @@ command-bearing opens; this does not add a human-confirm dialog or change mobile
long prompt shows only its end. So "holds the last send" accepts that end, each burst is two
Ctrl-Us per row on screen (`readClaudeDraftBox`) plus spares, and bursts repeat while what is left
ends earlier in the sent text (`endsEarlierIn`). Claude still writes such a prompt to the
- transcript as a branch it drops from its context, so the chat view keeps showing it (the reader
- does not follow `parentUuid`; known gap). The panel also makes the terminal's interrupt guess
+ transcript as a branch it drops from its context; the reader hides it once the re-sent prompt
+ is in the same read (`abandonedPromptUuids`: a childless prompt with a later non-tool-result
+ sibling under the same `parentUuid`). The panel also makes the terminal's interrupt guess
(`inferInterruptAfterSettle`), since an early Esc writes no transcript marker.
**An interrupt is drawn as Claude draws it, not as a user bubble.** Claude saves its marker
(`CLAUDE_INTERRUPT_MARKERS`, now in `shared/agents/claude-interrupt.ts`, shared with the context
diff --git a/src/core/chat-fixtures.test.ts b/src/core/chat-fixtures.test.ts
index b3820a734..3fc7c240d 100644
--- a/src/core/chat-fixtures.test.ts
+++ b/src/core/chat-fixtures.test.ts
@@ -632,8 +632,15 @@ function abandonedPrompts(): string {
// A later SIDECHAIN record with the same parent does not make a prompt abandoned.
/* 18 */ under(17, user('Check the lint.')),
/* 19 */ { ...under(17, user('subagent prompt')), isSidechain: true },
+ // Typed while a tool call is pending: the prompt and the call's (rejected) result share a
+ // parent and the answer hangs off the result. The model received the prompt, so it stays.
+ /* 20 */ under(18, assistant([toolUse('toolu_b1', 'Bash', { command: 'npm run lint' })])),
+ /* 21 */ under(20, user('Stop, run the tests instead.')),
+ /* 22 */ under(20, toolResult('toolu_b1', 'rejected')),
+ /* 23 */ under(22, user([text('[Request interrupted by user for tool use]')])),
+ /* 24 */ under(23, assistant([text('Running the tests.')])),
// The newest prompt has no answer yet and no sibling: shown.
- /* 20 */ under(18, user('One more thing.'))
+ /* 25 */ under(24, user('One more thing.'))
])
}
diff --git a/src/core/transcript-reader.abandoned.test.ts b/src/core/transcript-reader.abandoned.test.ts
index 929117893..d7b8516dd 100644
--- a/src/core/transcript-reader.abandoned.test.ts
+++ b/src/core/transcript-reader.abandoned.test.ts
@@ -60,6 +60,30 @@ describe('abandoned prompts', () => {
expect(texts(ms)).toEqual(['Explain the cache.', 'The cache keeps', 'Explain the cache briefly.'])
})
+ // Measured shape: a prompt typed while a tool call is pending interrupts the call. The prompt
+ // and the call's result share a parent, the answer hangs off the result's interrupt marker, and
+ // the model receives the prompt in that same request.
+ it('keeps a prompt typed during a pending tool call (its later sibling is the tool result)', () => {
+ const recs: Rec[] = [
+ prompt('u1', null, 'Run the lint.'),
+ {
+ type: 'assistant',
+ uuid: 'a1',
+ parentUuid: 'u1',
+ isSidechain: false,
+ message: { role: 'assistant', content: [{ type: 'tool_use', id: 't1', name: 'Bash', input: { command: 'npm run lint' } }] }
+ },
+ prompt('u2', 'a1', 'Stop, run the tests instead.'),
+ prompt('r1', 'a1', [{ type: 'tool_result', tool_use_id: 't1', content: 'rejected', is_error: true }]),
+ prompt('m1', 'r1', [{ type: 'text', text: '[Request interrupted by user for tool use]' }]),
+ reply('a2', 'm1', 'Running the tests.')
+ ]
+
+ const ms = parseChatMessages(recs.map((r) => JSON.stringify(r)))
+
+ expect(texts(ms)).toContain('Stop, run the tests instead.')
+ })
+
it('never hides a record without string uuid fields', () => {
const recs = queuedThenEsc.map((r) => ({ ...r, uuid: undefined }))
diff --git a/src/core/transcript-reader.ts b/src/core/transcript-reader.ts
index cc861a035..ea4230dd7 100644
--- a/src/core/transcript-reader.ts
+++ b/src/core/transcript-reader.ts
@@ -647,6 +647,10 @@ function isPromptContent(content: unknown): boolean {
return text
}
+function hasToolResult(content: unknown): boolean {
+ return Array.isArray(content) && content.some((c) => !!c && typeof c === 'object' && (c as { type?: unknown }).type === 'tool_result')
+}
+
/**
* The uuids of prompts the model never received. A transcript is a TREE: each record names its
* `parentUuid`, and the model is sent the chain from the newest record back to the root. When a
@@ -657,7 +661,10 @@ function isPromptContent(content: unknown): boolean {
* saw the text (5 were re-sent after an edit, one was a raw `/compact` recorded twice).
*
* A prompt is abandoned when no record in these lines has it as its parent AND a later
- * user/assistant record (not a sidechain) has the same parent. Both checks see only the lines
+ * user/assistant record (not a sidechain, not a tool result) has the same parent. The tool-result
+ * exclusion is measured (1 of 9 matches over 306 transcripts, 2,899 prompts): a prompt typed
+ * during a pending tool call is a childless sibling of that call's result, and the request that
+ * follows carries the prompt text. Both checks see only the lines
* given, so on a paged read a sibling in a NEWER window is missed and the prompt is shown, as
* before. A prompt the model did answer is safe: claude writes its first child right after it
* (measured in the 10 prompts that have both a child and a later sibling), so any window holding
@@ -670,7 +677,10 @@ function abandonedPromptUuids(records: readonly TreeFields[]): Set<string> {
records.forEach((o, i) => {
if (typeof o.parentUuid !== 'string') return
hasChild.add(o.parentUuid)
- if ((o.type === 'user' || o.type === 'assistant') && o.isSidechain !== true) {
+ // A tool-result record is never the re-send that marks its sibling abandoned: a prompt typed
+ // while a tool call is pending shares that call's parent with the result (the interrupt marker
+ // and the answer hang off the RESULT), and the model receives both in one request.
+ if ((o.type === 'user' || o.type === 'assistant') && o.isSidechain !== true && !hasToolResult(o.message?.content)) {
const list = byParent.get(o.parentUuid) ?? []
list.push({ i, uuid: o.uuid })
byParent.set(o.parentUuid, list)
diff --git a/src/shared/chat-fixtures/README.md b/src/shared/chat-fixtures/README.md
index 054364526..4b0ae2c09 100644
--- a/src/shared/chat-fixtures/README.md
+++ b/src/shared/chat-fixtures/README.md
@@ -89,7 +89,7 @@ the production paging, not a second implementation of it.
| `teammate-messages` | Rule **2b** of **System-injected records**: an idle notification with a report, one with no fields, a failed one (`failureReason`), a plain message whose summary is escaped (`&quot;`) and whose body quotes the close tag, a record holding two elements (a report, then a message), an unverified sender (`verified="false"`), the same text typed by a human (by `origin` and by `promptSource` alone — both stay user messages), the frame line with no element and an element with no `teammate_id` (both stay user messages). |
| `pasted-content` | **Pasted content**: a span between typed lines, backtick runs inside (a five-backtick fence) and the same id twice, an array text part beside an image, a pasted `<task-notification>` (stays the user's paste), a close tag with a different id inside a span (content), an empty paste, a same-id open nested inside a span (the span ends at the FIRST close), text that is not the CLI's grammar (no id, an upper-case or three- or five-digit id, a missing newline after the open or before the close: all left as typed) and an unclosed span (left as typed). |
| `queued-prompts` | **Queued prompts**: a typed prompt queued mid-turn (between a tool result and the reply), one with no `origin`, an array prompt with an image (its text block only), a paste inside one (fenced, as in a typed prompt), and six attachments that stay hidden: a task notification, a peer message (`isMeta`), a coordinator message, a non-human `origin`, a blank prompt and a non-`queued_command` attachment. |
-| `abandoned-prompts` | **Abandoned prompts**: a queued prompt cancelled with Esc (its parent is the interrupt marker) and a sent-then-edited array prompt are hidden; a raw `/compact` left beside its command record is hidden and the command stays; a prompt the model answered before a re-send, a prompt whose only later sibling is a sidechain record, and the newest unanswered prompt are all kept. |
+| `abandoned-prompts` | **Abandoned prompts**: a queued prompt cancelled with Esc (its parent is the interrupt marker) and a sent-then-edited array prompt are hidden; a raw `/compact` left beside its command record is hidden and the command stays; a prompt the model answered before a re-send, a prompt whose only later sibling is a sidechain record, a prompt typed during a pending tool call (its later sibling is the tool-result record), and the newest unanswered prompt are all kept. |
**`rich`** is a second desktop-only opt-in, beside `thinking`: the ⌘M panel reads with
`page.rich: true` (`ChatParseOptions.rich`) to get the parts only it draws. The fixtures never set
@@ -342,11 +342,19 @@ all of these hold:
`type:"text"` and a non-empty string `text` and no element with `type:"tool_result"`;
4. no record in the read has `parentUuid` equal to its `uuid` (any record type counts);
5. a LATER record in the read has the same `parentUuid`, `type` `"user"` or `"assistant"`,
- `isSidechain` not `true`, and a different `uuid`.
+ `isSidechain` not `true`, a different `uuid`, and a `message.content` that is NOT an array
+ holding an element with `type:"tool_result"`.
+
+The last clause of rule 5 is measured: a prompt typed while a tool call is pending interrupts the
+call, and Claude Code attaches the prompt and the call's tool-result record to the same parent (the
+interrupt marker and the answer hang off the result). That prompt has no child and a later sibling,
+yet the request that follows carries its text.
Both checks see only the read's records, so on a paged read a sibling in a NEWER window is not
seen and the prompt is shown, exactly as before this rule. Measured over 133 real transcripts
-(5,886 prompts): 6 records match, and the model received none of them.
+(5,886 prompts): 6 records match, and the model received none of them. A second set (306
+transcripts, 2,899 prompts) gave 9 matches before the tool-result clause, one of them the answered
+prompt above, and 8 after it.
## Renderer rules that follow from these records (desktop; iOS notes)
diff --git a/src/shared/chat-fixtures/expected/abandoned-prompts.pages.json b/src/shared/chat-fixtures/expected/abandoned-prompts.pages.json
index 4389d3533..3ef02beb3 100644
--- a/src/shared/chat-fixtures/expected/abandoned-prompts.pages.json
+++ b/src/shared/chat-fixtures/expected/abandoned-prompts.pages.json
@@ -164,6 +164,53 @@
"at": 1790362819000,
"key": 7220
},
+ {
+ "role": "assistant",
+ "parts": [
+ {
+ "kind": "tool",
+ "name": "Bash",
+ "arg": "npm run lint",
+ "id": "toolu_b1",
+ "result": "rejected"
+ }
+ ],
+ "at": 1790362820000,
+ "key": 7535
+ },
+ {
+ "role": "user",
+ "parts": [
+ {
+ "kind": "text",
+ "text": "Stop, run the tests instead."
+ }
+ ],
+ "at": 1790362821000,
+ "key": 8029
+ },
+ {
+ "role": "user",
+ "parts": [
+ {
+ "kind": "text",
+ "text": "[Request interrupted by user for tool use]"
+ }
+ ],
+ "at": 1790362823000,
+ "key": 8744
+ },
+ {
+ "role": "assistant",
+ "parts": [
+ {
+ "kind": "text",
+ "text": "Running the tests."
+ }
+ ],
+ "at": 1790362824000,
+ "key": 9112
+ },
{
"role": "user",
"parts": [
@@ -172,8 +219,8 @@
"text": "One more thing."
}
],
- "at": 1790362820000,
- "key": 7535
+ "at": 1790362825000,
+ "key": 9565
}
],
"olderCursor": null,
diff --git a/src/shared/chat-fixtures/inputs/abandoned-prompts.jsonl b/src/shared/chat-fixtures/inputs/abandoned-prompts.jsonl
index 08b9857f4..49f65f3f3 100644
--- a/src/shared/chat-fixtures/inputs/abandoned-prompts.jsonl
+++ b/src/shared/chat-fixtures/inputs/abandoned-prompts.jsonl
@@ -18,4 +18,9 @@
{"type":"user","message":{"role":"user","content":"<command-name>/compact</command-name>\n <command-message>compact</command-message>\n <command-args></command-args>"},"parentUuid":"00000000-0000-4000-8000-000000000016","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000017","timestamp":"2026-09-25T19:00:17.000Z"}
{"type":"user","message":{"role":"user","content":"Check the lint."},"parentUuid":"00000000-0000-4000-8000-000000000017","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000018","timestamp":"2026-09-25T19:00:18.000Z"}
{"type":"user","message":{"role":"user","content":"subagent prompt"},"parentUuid":"00000000-0000-4000-8000-000000000017","isSidechain":true,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000019","timestamp":"2026-09-25T19:00:19.000Z"}
-{"type":"user","message":{"role":"user","content":"One more thing."},"parentUuid":"00000000-0000-4000-8000-000000000018","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000020","timestamp":"2026-09-25T19:00:20.000Z"}
+{"type":"assistant","effort":"high","message":{"id":"msg_synthetic","type":"message","role":"assistant","model":"claude-opus-5-5","content":[{"type":"tool_use","id":"toolu_b1","name":"Bash","input":{"command":"npm run lint"}}],"stop_reason":null},"parentUuid":"00000000-0000-4000-8000-000000000018","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000020","timestamp":"2026-09-25T19:00:20.000Z"}
+{"type":"user","message":{"role":"user","content":"Stop, run the tests instead."},"parentUuid":"00000000-0000-4000-8000-000000000020","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000021","timestamp":"2026-09-25T19:00:21.000Z"}
+{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_b1","content":"rejected","is_error":false}]},"parentUuid":"00000000-0000-4000-8000-000000000020","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000022","timestamp":"2026-09-25T19:00:22.000Z"}
+{"type":"user","message":{"role":"user","content":[{"type":"text","text":"[Request interrupted by user for tool use]"}]},"parentUuid":"00000000-0000-4000-8000-000000000022","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000023","timestamp":"2026-09-25T19:00:23.000Z"}
+{"type":"assistant","effort":"high","message":{"id":"msg_synthetic","type":"message","role":"assistant","model":"claude-opus-5-5","content":[{"type":"text","text":"Running the tests."}],"stop_reason":null},"parentUuid":"00000000-0000-4000-8000-000000000023","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000024","timestamp":"2026-09-25T19:00:24.000Z"}
+{"type":"user","message":{"role":"user","content":"One more thing."},"parentUuid":"00000000-0000-4000-8000-000000000024","isSidechain":false,"sessionId":"5f0c2a9e-3b7d-4e61-9a24-7c1d8e0f4b32","cwd":"/srv/demo","version":"2.1.290","uuid":"00000000-0000-4000-8000-000000000025","timestamp":"2026-09-25T19:00:25.000Z"}