fix(session): prevent silent overflow on parent thread forks (#26912)

Lands #26912 from @markshields-tl with configurable session.parentForkMaxTokens and docs/tests/changelog updates.

Co-authored-by: Mark Shields <239231357+markshields-tl@users.noreply.github.com>
This commit is contained in:
Peter Steinberger
2026-02-25 23:53:43 +00:00
parent 8d1481cb4a
commit eb73e87f18
11 changed files with 211 additions and 14 deletions

View File

@@ -12,6 +12,7 @@ Docs: https://docs.openclaw.ai
### Fixes
- Slack/Session threads: prevent oversized parent-session inheritance from silently bricking new thread sessions, surface embedded context-overflow empty-result failures to users, and add configurable `session.parentForkMaxTokens` (default `100000`, `0` disables). (#26912) Thanks @markshields-tl.
- Security/Signal: enforce DM/group authorization before reaction-only notification enqueue so unauthorized senders can no longer inject Signal reaction system events under `dmPolicy`/`groupPolicy`; reaction notifications now require channel access checks first. This ships in the next npm release (`2026.2.25`). Thanks @tdjackey for reporting.
- Security/macOS beta onboarding: remove Anthropic OAuth sign-in and the legacy `oauth.json` onboarding path that exposed the PKCE verifier via OAuth `state`; this impacted the macOS beta onboarding path only. Anthropic subscription auth is now setup-token-only and will ship in the next npm release (`2026.2.25`). Thanks @zdi-disclosures for reporting.
- Security/Nextcloud Talk: drop replayed signed webhook events with persistent per-account replay dedupe across restarts, and reject unexpected webhook backend origins when account base URL is configured. Thanks @aristorechina for reporting.

View File

@@ -1250,6 +1250,7 @@ See [Multi-Agent Sandbox & Tools](/tools/multi-agent-sandbox-tools) for preceden
},
resetTriggers: ["/new", "/reset"],
store: "~/.openclaw/agents/{agentId}/sessions/sessions.json",
parentForkMaxTokens: 100000, // skip parent-thread fork above this token count (0 disables)
maintenance: {
mode: "warn", // warn | enforce
pruneAfter: "30d",
@@ -1283,6 +1284,9 @@ See [Multi-Agent Sandbox & Tools](/tools/multi-agent-sandbox-tools) for preceden
- **`identityLinks`**: map canonical ids to provider-prefixed peers for cross-channel session sharing.
- **`reset`**: primary reset policy. `daily` resets at `atHour` local time; `idle` resets after `idleMinutes`. When both configured, whichever expires first wins.
- **`resetByType`**: per-type overrides (`direct`, `group`, `thread`). Legacy `dm` accepted as alias for `direct`.
- **`parentForkMaxTokens`**: max parent-session `totalTokens` allowed when creating a forked thread session (default `100000`).
- If parent `totalTokens` is above this value, OpenClaw starts a fresh thread session instead of inheriting parent transcript history.
- Set `0` to disable this guard and always allow parent forking.
- **`mainKey`**: legacy field. Runtime now always uses `"main"` for the main direct-chat bucket.
- **`sendPolicy`**: match by `channel`, `chatType` (`direct|group|channel`, with legacy `dm` alias), `keyPrefix`, or `rawKeyPrefix`. First deny wins.
- **`maintenance`**: session-store cleanup + retention controls.

View File

@@ -128,6 +128,7 @@ Rules of thumb:
- **Reset** (`/new`, `/reset`) creates a new `sessionId` for that `sessionKey`.
- **Daily reset** (default 4:00 AM local time on the gateway host) creates a new `sessionId` on the next message after the reset boundary.
- **Idle expiry** (`session.reset.idleMinutes` or legacy `session.idleMinutes`) creates a new `sessionId` when a message arrives after the idle window. When daily + idle are both configured, whichever expires first wins.
- **Thread parent fork guard** (`session.parentForkMaxTokens`, default `100000`) skips parent transcript forking when the parent session is already too large; the new thread starts fresh. Set `0` to disable.
Implementation detail: the decision happens in `initSessionState()` in `src/auto-reply/reply/session.ts`.

View File

@@ -572,6 +572,22 @@ export async function runAgentTurnWithFallback(params: {
}
}
// If the run completed but with an embedded context overflow error that
// wasn't recovered from (e.g. compaction reset already attempted), surface
// the error to the user instead of silently returning an empty response.
// See #26905: Slack DM sessions silently swallowed messages when context
// overflow errors were returned as embedded error payloads.
const finalEmbeddedError = runResult?.meta?.error;
const hasPayloadText = runResult?.payloads?.some((p) => p.text?.trim());
if (finalEmbeddedError && isContextOverflowError(finalEmbeddedError.message) && !hasPayloadText) {
return {
kind: "final",
payload: {
text: "⚠️ Context overflow — this conversation is too large for the model. Use /new to start a fresh session.",
},
};
}
return {
kind: "success",
runId,

View File

@@ -205,6 +205,130 @@ describe("initSessionState thread forking", () => {
warn.mockRestore();
});
it("skips fork and creates fresh session when parent tokens exceed threshold", async () => {
const root = await makeCaseDir("openclaw-thread-session-overflow-");
const sessionsDir = path.join(root, "sessions");
await fs.mkdir(sessionsDir);
const parentSessionId = "parent-overflow";
const parentSessionFile = path.join(sessionsDir, "parent.jsonl");
const header = {
type: "session",
version: 3,
id: parentSessionId,
timestamp: new Date().toISOString(),
cwd: process.cwd(),
};
const message = {
type: "message",
id: "m1",
parentId: null,
timestamp: new Date().toISOString(),
message: { role: "user", content: "Parent prompt" },
};
await fs.writeFile(
parentSessionFile,
`${JSON.stringify(header)}\n${JSON.stringify(message)}\n`,
"utf-8",
);
const storePath = path.join(root, "sessions.json");
const parentSessionKey = "agent:main:slack:channel:c1";
// Set totalTokens well above PARENT_FORK_MAX_TOKENS (100_000)
await saveSessionStore(storePath, {
[parentSessionKey]: {
sessionId: parentSessionId,
sessionFile: parentSessionFile,
updatedAt: Date.now(),
totalTokens: 170_000,
},
});
const cfg = {
session: { store: storePath },
} as OpenClawConfig;
const threadSessionKey = "agent:main:slack:channel:c1:thread:456";
const result = await initSessionState({
ctx: {
Body: "Thread reply",
SessionKey: threadSessionKey,
ParentSessionKey: parentSessionKey,
},
cfg,
commandAuthorized: true,
});
// Should be marked as forked (to prevent re-attempts) but NOT actually forked from parent
expect(result.sessionEntry.forkedFromParent).toBe(true);
// Session ID should NOT match the parent — it should be a fresh UUID
expect(result.sessionEntry.sessionId).not.toBe(parentSessionId);
// Session file should NOT be the parent's file (it was not forked)
expect(result.sessionEntry.sessionFile).not.toBe(parentSessionFile);
});
it("respects session.parentForkMaxTokens override", async () => {
const root = await makeCaseDir("openclaw-thread-session-overflow-override-");
const sessionsDir = path.join(root, "sessions");
await fs.mkdir(sessionsDir);
const parentSessionId = "parent-override";
const parentSessionFile = path.join(sessionsDir, "parent.jsonl");
const header = {
type: "session",
version: 3,
id: parentSessionId,
timestamp: new Date().toISOString(),
cwd: process.cwd(),
};
const message = {
type: "message",
id: "m1",
parentId: null,
timestamp: new Date().toISOString(),
message: { role: "user", content: "Parent prompt" },
};
await fs.writeFile(
parentSessionFile,
`${JSON.stringify(header)}\n${JSON.stringify(message)}\n`,
"utf-8",
);
const storePath = path.join(root, "sessions.json");
const parentSessionKey = "agent:main:slack:channel:c1";
await saveSessionStore(storePath, {
[parentSessionKey]: {
sessionId: parentSessionId,
sessionFile: parentSessionFile,
updatedAt: Date.now(),
totalTokens: 170_000,
},
});
const cfg = {
session: {
store: storePath,
parentForkMaxTokens: 200_000,
},
} as OpenClawConfig;
const threadSessionKey = "agent:main:slack:channel:c1:thread:789";
const result = await initSessionState({
ctx: {
Body: "Thread reply",
SessionKey: threadSessionKey,
ParentSessionKey: parentSessionKey,
},
cfg,
commandAuthorized: true,
});
expect(result.sessionEntry.forkedFromParent).toBe(true);
expect(result.sessionEntry.sessionFile).toBeTruthy();
const forkedContent = await fs.readFile(result.sessionEntry.sessionFile ?? "", "utf-8");
expect(forkedContent).toContain(parentSessionFile);
});
it("records topic-specific session files when MessageThreadId is present", async () => {
const root = await makeCaseDir("openclaw-topic-session-");
const storePath = path.join(root, "sessions.json");

View File

@@ -105,6 +105,21 @@ export type SessionInitResult = {
triggerBodyNormalized: string;
};
/**
* Default max parent token count beyond which thread/session parent forking is skipped.
* This prevents new thread sessions from inheriting near-full parent context.
* See #26905.
*/
const DEFAULT_PARENT_FORK_MAX_TOKENS = 100_000;
function resolveParentForkMaxTokens(cfg: OpenClawConfig): number {
const configured = cfg.session?.parentForkMaxTokens;
if (typeof configured === "number" && Number.isFinite(configured) && configured >= 0) {
return Math.floor(configured);
}
return DEFAULT_PARENT_FORK_MAX_TOKENS;
}
function forkSessionFromParent(params: {
parentEntry: SessionEntry;
agentId: string;
@@ -171,6 +186,7 @@ export async function initSessionState(params: {
const resetTriggers = sessionCfg?.resetTriggers?.length
? sessionCfg.resetTriggers
: DEFAULT_RESET_TRIGGERS;
const parentForkMaxTokens = resolveParentForkMaxTokens(cfg);
const sessionScope = sessionCfg?.scope ?? "per-sender";
const storePath = resolveStorePath(sessionCfg?.store, { agentId });
@@ -399,21 +415,33 @@ export async function initSessionState(params: {
sessionStore[parentSessionKey] &&
!alreadyForked
) {
log.warn(
`forking from parent session: parentKey=${parentSessionKey} → sessionKey=${sessionKey} ` +
`parentTokens=${sessionStore[parentSessionKey].totalTokens ?? "?"}`,
);
const forked = forkSessionFromParent({
parentEntry: sessionStore[parentSessionKey],
agentId,
sessionsDir: path.dirname(storePath),
});
if (forked) {
sessionId = forked.sessionId;
sessionEntry.sessionId = forked.sessionId;
sessionEntry.sessionFile = forked.sessionFile;
const parentTokens = sessionStore[parentSessionKey].totalTokens ?? 0;
if (parentForkMaxTokens > 0 && parentTokens > parentForkMaxTokens) {
// Parent context is too large — forking would create a thread session
// that immediately overflows the model's context window. Start fresh
// instead and mark as forked to prevent re-attempts. See #26905.
log.warn(
`skipping parent fork (parent too large): parentKey=${parentSessionKey} → sessionKey=${sessionKey} ` +
`parentTokens=${parentTokens} maxTokens=${parentForkMaxTokens}`,
);
sessionEntry.forkedFromParent = true;
log.warn(`forked session created: file=${forked.sessionFile}`);
} else {
log.warn(
`forking from parent session: parentKey=${parentSessionKey} → sessionKey=${sessionKey} ` +
`parentTokens=${parentTokens}`,
);
const forked = forkSessionFromParent({
parentEntry: sessionStore[parentSessionKey],
agentId,
sessionsDir: path.dirname(storePath),
});
if (forked) {
sessionId = forked.sessionId;
sessionEntry.sessionId = forked.sessionId;
sessionEntry.sessionFile = forked.sessionFile;
sessionEntry.forkedFromParent = true;
log.warn(`forked session created: file=${forked.sessionFile}`);
}
}
}
const fallbackSessionFile = !sessionEntry.sessionFile

View File

@@ -973,6 +973,8 @@ export const FIELD_HELP: Record<string, string> = {
"Controls interval for repeated typing indicators while replies are being prepared in typing-capable channels. Increase to reduce chatty updates or decrease for more active typing feedback.",
"session.typingMode":
'Controls typing behavior timing: "never", "instant", "thinking", or "message" based emission points. Keep conservative modes in high-volume channels to avoid unnecessary typing noise.',
"session.parentForkMaxTokens":
"Maximum parent-session token count allowed for thread/session inheritance forking. If the parent exceeds this, OpenClaw starts a fresh thread session instead of forking; set 0 to disable this protection.",
"session.mainKey":
'Overrides the canonical main session key used for continuity when dmScope or routing logic points to "main". Use a stable value only if you intentionally need custom session anchoring.',
"session.sendPolicy":

View File

@@ -455,6 +455,7 @@ export const FIELD_LABELS: Record<string, string> = {
"session.store": "Session Store Path",
"session.typingIntervalSeconds": "Session Typing Interval (seconds)",
"session.typingMode": "Session Typing Mode",
"session.parentForkMaxTokens": "Session Parent Fork Max Tokens",
"session.mainKey": "Session Main Key",
"session.sendPolicy": "Session Send Policy",
"session.sendPolicy.default": "Session Send Policy Default Action",

View File

@@ -112,6 +112,12 @@ export type SessionConfig = {
store?: string;
typingIntervalSeconds?: number;
typingMode?: TypingMode;
/**
* Max parent transcript token count allowed for thread/session forking.
* If parent totalTokens is above this value, OpenClaw skips parent fork and
* starts a fresh thread session instead. Set to 0 to disable this guard.
*/
parentForkMaxTokens?: number;
mainKey?: string;
sendPolicy?: SessionSendPolicyConfig;
agentToAgent?: {

View File

@@ -14,6 +14,19 @@ describe("SessionSchema maintenance extensions", () => {
).not.toThrow();
});
it("accepts parentForkMaxTokens including 0 to disable the guard", () => {
expect(() => SessionSchema.parse({ parentForkMaxTokens: 100_000 })).not.toThrow();
expect(() => SessionSchema.parse({ parentForkMaxTokens: 0 })).not.toThrow();
});
it("rejects negative parentForkMaxTokens", () => {
expect(() =>
SessionSchema.parse({
parentForkMaxTokens: -1,
}),
).toThrow(/parentForkMaxTokens/i);
});
it("accepts disabling reset archive cleanup", () => {
expect(() =>
SessionSchema.parse({

View File

@@ -52,6 +52,7 @@ export const SessionSchema = z
store: z.string().optional(),
typingIntervalSeconds: z.number().int().positive().optional(),
typingMode: TypingModeSchema.optional(),
parentForkMaxTokens: z.number().int().nonnegative().optional(),
mainKey: z.string().optional(),
sendPolicy: SessionSendPolicySchema.optional(),
agentToAgent: z