fix(coding-agent): guard against stale kept pre-compaction usage in error-path threshold check

This commit is contained in:
Mario Zechner
2026-03-06 16:57:06 +01:00
parent c950c692a1
commit d1a17bbae5
2 changed files with 95 additions and 1 deletions

View File

@@ -1736,8 +1736,20 @@ export class AgentSession {
// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.
let contextTokens: number;
if (assistantMessage.stopReason === "error") {
const estimate = estimateContextTokens(this.agent.state.messages);
const messages = this.agent.state.messages;
const estimate = estimateContextTokens(messages);
if (estimate.lastUsageIndex === null) return; // No usage data at all
// Verify the usage source is post-compaction. Kept pre-compaction messages
// have stale usage reflecting the old (larger) context and would falsely
// trigger compaction right after one just finished.
const usageMsg = messages[estimate.lastUsageIndex];
if (
compactionEntry &&
usageMsg.role === "assistant" &&
(usageMsg as AssistantMessage).timestamp <= new Date(compactionEntry.timestamp).getTime()
) {
return;
}
contextTokens = estimate.tokens;
} else {
contextTokens = calculateContextTokens(assistantMessage.usage);

View File

@@ -351,4 +351,86 @@ describe("AgentSession auto-compaction queue resume", () => {
expect(runAutoCompactionSpy).not.toHaveBeenCalled();
});
it("should not trigger threshold compaction for error messages when only kept pre-compaction usage exists", async () => {
const model = session.model!;
const preCompactionTimestamp = Date.now() - 10_000;
// A "kept" assistant message from before compaction with high usage
const keptAssistant: AssistantMessage = {
role: "assistant",
content: [{ type: "text", text: "kept response from before compaction" }],
api: model.api,
provider: model.provider,
model: model.id,
usage: {
input: 180_000,
output: 10_000,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 190_000,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "stop",
timestamp: preCompactionTimestamp,
};
// Record the kept assistant in the session and create a compaction after it
sessionManager.appendMessage({
role: "user",
content: [{ type: "text", text: "before compaction" }],
timestamp: preCompactionTimestamp - 1000,
});
sessionManager.appendMessage(keptAssistant);
const firstKeptEntryId = sessionManager.getEntries()[0]!.id;
sessionManager.appendCompaction("summary", firstKeptEntryId, keptAssistant.usage.totalTokens, undefined, false);
// Post-compaction error message
const errorAssistant: AssistantMessage = {
role: "assistant",
content: [{ type: "text", text: "" }],
api: model.api,
provider: model.provider,
model: model.id,
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
totalTokens: 0,
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
},
stopReason: "error",
errorMessage: "529 overloaded",
timestamp: Date.now(),
};
// Agent state has the kept assistant (pre-compaction) and the error (post-compaction)
session.agent.replaceMessages([
{ role: "user", content: [{ type: "text", text: "kept user msg" }], timestamp: preCompactionTimestamp - 1000 },
keptAssistant,
{ role: "user", content: [{ type: "text", text: "new prompt" }], timestamp: Date.now() - 500 },
errorAssistant,
]);
const runAutoCompactionSpy = vi
.spyOn(
session as unknown as {
_runAutoCompaction: (reason: "overflow" | "threshold", willRetry: boolean) => Promise<void>;
},
"_runAutoCompaction",
)
.mockResolvedValue();
const checkCompaction = (
session as unknown as {
_checkCompaction: (assistantMessage: AssistantMessage, skipAbortedCheck?: boolean) => Promise<void>;
}
)._checkCompaction.bind(session);
await checkCompaction(errorAssistant);
// Should NOT compact because the only usage data is from a kept pre-compaction message
expect(runAutoCompactionSpy).not.toHaveBeenCalled();
});
});