hansjone/dsh-im-ops
1
0
Fork 0
mirror of https://github.com/hansjone/dsh-im-ops.git synced 2026-10-11 01:43:25 +08:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix(feishu): renew the reply wait window while the turn is active

Browse source 
A long-running Harness task was falsely reported as MODEL_REPLY_TIMEOUT
after the fixed 10-minute deadline even though the backend was still
working. Replace the hard wall-clock deadline with an activity-based
stall window: the wait renews whenever the turn produces new events, the
control ownership is active, or (mid-turn with no new events) Harness
still reports the session as running. Only a genuine stall with no
progress for the whole timeoutMs window fails fast.

This is the shared harness-client layer, exercised end-to-end by the
Feishu channel (which already exposes HARNESS_REPLY_TIMEOUT_MS); other
channels inherit the same fix.

Adds two regression tests: a long-running turn with periodic activity
completes instead of timing out, and a genuinely stalled turn still
fails with harness-reply-timeout.
This commit is contained in:
C3H3-AI 2026-09-01 22:24:33 +08:00
parent c2be2389b0
commit 955c2217fd
3 changed files with 132 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

34
src/channels/shared/harness-client.mjs
View file

@ -428,6 +428,11 @@ export class HarnessReplyTracker {
return this.#finished;
}
/** The highest event seq consumed so far; advances as the turn produces events. */
get lastSeq() {
return this.#lastSeq;
}
get answer() {
return this.#latestText.trim();
}
@ -1387,8 +1392,13 @@ export class HarnessClient {
promptAccepted = true;
try {
const deadline = Date.now() + timeoutMs;
while (Date.now() < deadline) {
// Renew the wait window whenever the turn is still making progress, so
// a long-running task is not falsely reported as MODEL_REPLY_TIMEOUT
// while Harness is still working. A genuine stall (no new events and
// no active turn for the whole timeoutMs window) still fails fast.
let lastProgressAt = Date.now();
let lastPollSeq = tracker.lastSeq;
while (Date.now() - lastProgressAt < timeoutMs) {
await sleep(300, signal);
const history = await this.rpc(
'session.history',
@ -1402,6 +1412,26 @@ export class HarnessClient {
if (!wasActive && ownership.active) ownership.reconnect?.();
}
const updates = tracker.consumeAll(history.events ?? []);
// Progress = the turn produced new events, or the ownership is still
// active. Either refreshes the stall window.
const seqAdvanced = tracker.lastSeq > lastPollSeq;
lastPollSeq = tracker.lastSeq;
if (seqAdvanced) {
lastProgressAt = Date.now();
} else if (ownership?.active) {
lastProgressAt = Date.now();
} else if (tracker.tracking) {
// Mid-turn but no new events: ask Harness whether the session is
// still running before declaring a stall. A failing probe must not
// mask a real stall, so it is not treated as progress.
try {
if (await this.isSessionRunning(sessionId, { signal })) {
lastProgressAt = Date.now();
}
} catch {
// ignore probe failure; the deadline still applies
}
}
if (onUpdate) {
const visibleUpdates = progressMode === 'all' ? updates : updates.slice(-1);
for (const update of visibleUpdates) {