Skip to content

Fix Claude Code follow-up/compact 400s and duplicated assistant text #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
zhaohui-yang wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix Claude Code follow-up/compact 400s and duplicated assistant text #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 25 additions & 13 deletions src/anthropic.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,16 +93,31 @@ export function normalizeTools(tools) {
* Dig tool results out of this batch: Anthropic nests `tool_result` blocks in
* user messages (no standalone `role:"tool"`), one level deeper than OpenAI.
*
* Returns null when the batch carries no tool results at all, `{turn,
* Returns null when the trailing turn carries no tool results, `{turn,
* results, orphan: null}` when at least one id matches a live turn, and
* `{turn: null, results, orphan: [ids]}` when results exist but no id
* matches anything. The caller must reject the orphan case — resuming
* nothing would silently bill a second run for the same turn.
* `{turn: null, results, orphan: [ids]}` when trailing results exist but
* no id is live. The caller starts a new run for the orphan case: Claude
* Code compact/follow-up resends finished tool_result blocks, and a 400
* aborts autocompact.
*/
function messagesAfterLastAssistant(messages) {
let start = 0;
for (let i = (messages ?? []).length - 1; i >= 0; i--) {
if (messages[i]?.role === "assistant") {
start = i + 1;
break;
}
}
return (messages ?? []).slice(start);
}

function digToolResults(messages) {
const results = [];
let turn = null;
for (const m of messages ?? []) {
// Only the trailing user turn after the last assistant reply can resume a
// live RelayTurn. Historical tool_result blocks in earlier user messages
// belong to finished turns; treating them as resume 400s the next prompt.
for (const m of messagesAfterLastAssistant(messages)) {
if (m?.role !== "user" || !Array.isArray(m.content)) continue;
for (const block of m.content) {
if (block?.type !== "tool_result" || !block.tool_use_id) continue;
Expand Down Expand Up @@ -323,14 +338,11 @@ const adapter = {

const resume = digToolResults(messages);
if (resume && !resume.turn) {
// Tool results with no live turn behind them: the original turn expired
// (result timeout, server restart, client resend). Resuming would feed
// nothing; treating the batch as fresh would bill a second run for the
// same turn. Say so and let the client retry.
return {
error: `no matching pending tool call (${resume.orphan.join(", ")}); the turn may have expired`,
status: 400,
};
// Stale tool_result after the last assistant (compaction, follow-up, or
// an already-finished turn). 400 here aborts Claude Code auto-compact
// and surfaces as "Prompt is too long · automatic compaction failed".
// Start a new run with the folded prompt instead of resume.
log.warn(`stale tool_result id(s) ${resume.orphan.join(", ")}; starting a new run`);
}

return {
Expand Down
29 changes: 17 additions & 12 deletions src/openai.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,15 +37,26 @@ export function normalizeTools(tools) {

/**
* Dig tool results out of this batch: `role:"tool"` messages keyed by
* tool_call_id. Returns null with no tool results, `{turn, results}` when an
* id matches a live turn, `{orphan: [ids]}` when results exist but nothing
* matches — the caller must reject the orphan case, or the batch would
* silently bill a second run for a turn that no longer exists.
* tool_call_id. Returns null with no trailing tool results, `{turn, results}`
* when an id matches a live turn, `{orphan: [ids]}` when trailing results
* exist but nothing is live. The caller starts a new run for the orphan
* case so Claude Code compact/follow-up is not aborted with 400.
*/
function messagesAfterLastAssistant(messages) {
let start = 0;
for (let i = (messages ?? []).length - 1; i >= 0; i--) {
if (messages[i]?.role === "assistant") {
start = i + 1;
break;
}
}
return (messages ?? []).slice(start);
}

function digToolResults(messages) {
const results = [];
let turn = null;
for (const m of messages) {
for (const m of messagesAfterLastAssistant(messages)) {
if (m?.role !== "tool" || !m.tool_call_id) continue;
results.push({ id: m.tool_call_id, content: m.content, isError: m.is_error === true });
turn = turn ?? lookupTurn(m.tool_call_id);
Expand All @@ -63,13 +74,7 @@ const adapter = {

const resume = digToolResults(messages);
if (resume?.orphan) {
// Tool results with no live turn behind them (expired timer, server
// restart, client resend): resuming would feed nothing, and treating
// the batch as fresh would bill a second run for the same turn.
return {
error: `no matching pending tool call (${resume.orphan.join(", ")}); the turn may have expired`,
status: 400,
};
log.warn(`stale tool_result id(s) ${resume.orphan.join(", ")}; starting a new run`);
}

return {
Expand Down
11 changes: 10 additions & 1 deletion src/tool-relay.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,6 +177,8 @@ export class RelayTurn {
this.callIdPrefix = "call_";
/** Calls arriving while the sink is closed wait here for the next attach to replay them. */
this.parked = [];
/** True once any assistant text has been handed to the sink or pendingText. */
this.sawText = false;
}

#touch() {
Expand Down Expand Up @@ -251,6 +253,7 @@ export class RelayTurn {
#emitText(t) {
if (!t) return;
this.#touch();
this.sawText = true;
if (this.sink) this.sink.text(t);
else this.pendingText += t;
}
Expand Down Expand Up @@ -390,7 +393,13 @@ export class RelayTurn {
}

get #hasText() {
return this.pendingText.length > 0 || (this.sink?.parts?.length ?? 0) > 0;
// AnthropicSseWriter stores text in `accum`, not `parts`. After a streamed
// assistant event, pendingText is empty and sink.parts is undefined, so
// the old check treated the turn as empty and appended result.result —
// the client then saw the same paragraph twice.
return this.sawText || this.pendingText.length > 0
|| (this.sink?.parts?.length ?? 0) > 0
|| (this.sink?.accum?.length ?? 0) > 0;
}
}

Expand Down
45 changes: 31 additions & 14 deletions test-protocol.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -348,10 +348,11 @@ test("collect sink: converts usage, keeps content", () => {
assert.equal(c.content, "answer");
});

// ── entry guard: orphan tool results must 400, never bill a new run ────
// A tool_result whose id matches no live turn used to be treated as a fresh
// request: new account, new run, second billing for the same turn. It must
// come back as 400 invalid_request_error so the client retries instead.
// ── stale tool_result: new run, not 400 ────
// Claude Code resends finished tool_result blocks on compact/follow-up.
// A 400 aborts autocompact ("Prompt is too long · compaction failed").
// Empty test pool still answers 502 before any run; anything but 400
// proves the orphan guard no longer rejects the request.
function fakeHttpRes() {
return {
status: 0,
Expand All @@ -362,30 +363,46 @@ function fakeHttpRes() {
};
}

test("entry: anthropic tool results with no matching turn -> 400 invalid_request_error", async () => {
test("entry: anthropic tool results with no matching turn start a new run, not 400", async () => {
const res = fakeHttpRes();
await A.handleMessages({
model: "claude-opus-5",
messages: [
{ role: "user", content: [{ type: "tool_result", tool_use_id: "toolu_deadbeef", content: "x" }] },
],
}, res);
assert.equal(res.status, 400);
const body = JSON.parse(res.body);
assert.equal(body.error.type, "invalid_request_error");
assert.match(body.error.message, /no matching pending tool call/);
assert.notEqual(res.status, 400, "stale tool_result must not abort compact/follow-up");
assert.equal(res.status, 502, "empty pool answers 502 before any run");
assert.ok(
recentLogs(50).some((e) => e.level === "warn" && /stale tool_result/.test(e.msg)),
"the fallback must be logged",
);
});

test("entry: openai tool results with no matching turn -> 400 invalid_request_error", async () => {
test("entry: openai tool results with no matching turn start a new run, not 400", async () => {
const res = fakeHttpRes();
await handleChat({
model: "x",
messages: [{ role: "tool", tool_call_id: "call_deadbeef", content: "x" }],
}, res);
assert.equal(res.status, 400);
const body = JSON.parse(res.body);
assert.equal(body.error.type, "invalid_request_error");
assert.match(body.error.message, /no matching pending tool call/);
assert.notEqual(res.status, 400, "stale tool_result must not abort compact/follow-up");
assert.equal(res.status, 502, "empty pool answers 502 before any run");
});

test("entry: historical tool_result before the last assistant does not look like a resume", async () => {
const res = fakeHttpRes();
await A.handleMessages({
model: "claude-opus-5",
messages: [
{ role: "user", content: "search this" },
{ role: "assistant", content: [{ type: "tool_use", id: "toolu_old", name: "WebSearch", input: {} }] },
{ role: "user", content: [{ type: "tool_result", tool_use_id: "toolu_old", content: "done" }] },
{ role: "assistant", content: "here is the answer" },
{ role: "user", content: "follow-up question" },
],
}, res);
assert.notEqual(res.status, 400, "finished tool rounds in history must not 400");
assert.equal(res.status, 502, "empty pool answers 502 before any run");
});

test("entry: a fresh request without tool results never trips the orphan guard", async () => {
Expand Down