stop resume-cursor guessing that a session is claude

B7. `msg.model || DEFAULT_MODEL` declared every session without an explicit model to be
claude-code, and the parity doc recorded only the visible half of what that cost.

The durable false cut-off is real: endTurnIfAgentIsGone asked the claude sidecar about a
key it had never held, was told false, and wrote "the agent went away" into a turn that
was running fine. It survives reload, because surviving reload is what that row is for.

The same default also handed the session to adoptOrphanedSession as a claude one, which
subscribes it to that sidecar bus and pins session.model — so an opencode turn output
never arrived, and stopping it called killClaude on a key that sidecar never had. A stop
button that silently does nothing.

decideResume makes both rules explicit: the server record beats the client claim, and an
unknown harness stays unknown — no adoption, no cut-off check, just the replay. Silence
is the safe failure when the wrong answer is written durably.

DEFAULT_MODEL stays in handleAttach and is now commented as to why: that path reached its
sessionId by asking the claude sidecar to resolve a claudeSessionId, so only claude could
have answered.

First test in api/chat, which had none. websocket.ts has no seam to drive the handler
through, so the decision is extracted and tested; the wiring around it is not covered.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-10 13:16:30 +01:00
co-authored by Claude Opus 5
parent be259f813a
commit d9a3513cb5
2 changed files with 99 additions and 5 deletions
+55 -5
View File
@@ -614,11 +614,22 @@ async function handleResumeCursor(
msg: { sessionId: string; cursor: number; model?: string; cwd?: string; generating?: boolean },
): Promise<void> {
const { sessionId, cursor } = msg;
const model = msg.model || DEFAULT_MODEL;
if (!sessionManager.getSession(sessionId)) {
adoptOrphanedSession(ws, sessionId, model, msg.cwd ?? '');
const known = sessionManager.getSession(sessionId);
const decision = decideResume(known?.model, msg.model);
const model = decision.kind === 'replay-only' ? null : decision.model;
if (decision.kind === 'adopt') {
adoptOrphanedSession(ws, sessionId, decision.model, msg.cwd ?? '');
} else if (decision.kind === 'replay-only') {
// Replay the durable log and stop there. Not adopting costs a live re-subscription; adopting on a
// guess cost correctness — see decideResume.
logger.warn('resume-cursor names a session this process does not know, and no model; replaying only', {
sessionId,
});
}
sessionManager.attachWs(sessionId, ws);
sessionManager.attachWs(sessionId, ws); // a no-op when adoption was skipped
wsToSessionMap.set(ws as any, sessionId);
try {
const events = await getChatEventsSince(sessionId, cursor ?? 0);
@@ -628,7 +639,42 @@ async function handleResumeCursor(
} catch (err) {
logger.error('Failed to replay chat events on resume-cursor', { sessionId, error: String(err) });
}
if (msg.generating) await endTurnIfAgentIsGone(ws, sessionId, model);
// Only ask when the harness is actually known. The check writes a DURABLE row, so a wrong answer here
// is permanent — silence is the safe failure.
if (msg.generating && model) await endTurnIfAgentIsGone(ws, sessionId, model);
}
export type ResumeDecision =
| { kind: 'known'; model: string }
| { kind: 'adopt'; model: string }
| { kind: 'replay-only' };
/**
* Which harness a resuming session belongs to, and what that permits.
*
* This used to be `msg.model || DEFAULT_MODEL`, and `DEFAULT_MODEL` is `claude-code`. So a resume-cursor
* that omitted `model` declared every session — OpenCode ones included — to be Claude, with three
* consequences that all read as something else:
*
* - **Adopted into the wrong harness.** `adoptOrphanedSession` subscribes to the sidecar bus for that
* model and pins `session.model` for the rest of its life, so an OpenCode turn's output never
* arrived, and stopping it called `killClaude` on a key that sidecar had never held — a dead stop
* button, silently.
* - **A durable false `cut-off`.** `endTurnIfAgentIsGone` asked the Claude sidecar whether it was
* generating, was told `false` because it had never heard of the session, and wrote "the agent went
* away" into a turn that was running normally. It survives reload, which is the whole point of
* writing it durably, and is therefore unrecoverable from the UI.
* - Masked, never fixed, by the client always happening to send `model` next to `sessionId`.
*
* Two rules. **The server's own record beats the client's claim** — a session in memory already knows its
* harness, and letting a socket re-declare it is how the wrong sidecar gets a session in the first place.
* **An unknown harness stays unknown**: no adoption, no cut-off check, just the replay. Defaulting is
* what made a guess indistinguishable from knowledge.
*/
export function decideResume(knownModel: string | undefined, claimedModel: string | undefined): ResumeDecision {
if (knownModel) return { kind: 'known', model: knownModel };
if (claimedModel) return { kind: 'adopt', model: claimedModel };
return { kind: 'replay-only' };
}
/**
@@ -683,6 +729,10 @@ async function handleAttach(ws: ServerWebSocket<WSData>, msg: { claudeSessionId:
// An officer restart takes the in-memory session with it while the agent carries on, so the key can
// resolve to a session this process has never heard of. Adopting re-subscribes it to the sidecar's bus,
// which is what makes the rest of the turn arrive.
// `DEFAULT_MODEL` is right here and wrong in resume-cursor, which is worth being explicit about since
// it was just removed there: this path reached `sessionId` by asking the CLAUDE sidecar to resolve a
// `claudeSessionId`, so the harness is not a guess — only Claude could have answered. Attach is
// Claude-only by construction. If an OpenCode reattach verb ever lands, this stops being safe.
const existing = sessionManager.getSession(sessionId);
const session = existing ?? adoptOrphanedSession(ws, sessionId, DEFAULT_MODEL, '');