Files
PaperClipAI/server/src/services/native-runtime/remote-warm-session-recovery.ts
T
Nicky LeachandClaude Opus 5.5 0173110ee3 fix(native): point failed run attachments at the docs (#19052)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - When a later run cannot attach to a native session, the run fails,
and a later run may start a new provider thread.
> - The first failure is easy to miss. In 5 of 11 cases on a test
deployment, the run ended with a different error before it reported the
attach failure.
> - People and agents who diagnose a failed run look at its log and
events, so that is where an explanation is useful.
> - This pull request adds a pointer there at the first failure, with
the real cause and a link to the docs.
> - The benefit is that a failed attach explains itself, and readers can
find the open follow-up in #19047.

## Linked Issues or Issue Description

Refs #19047. I searched open and closed pull requests and issues for
`prior_attach_failed` and related terms, and found no duplicate.

## What Changed

- **Log:** `executeHeartbeatRuntime` writes one stderr line to the run
log when a same-run retry follows a failed `run.attach`.
- **Events:** `heartbeat.ts` adds a `lifecycle` event with the same
message and the code `native_run_attach_failed`, using the stored
redacted failure.
- **Helper:** `describeFailedAttachment` extracts the real cause and
returns nothing for other failures.
- **Docs:** "Session settlement after a run" in
`doc/execution-semantics.md` now explains this and links #19047.
- **Link target:** the pointer links to the repo docs and not the issue.
Agents read these logs, and anyone can comment on a public issue.
- **Unchanged:** error codes and retry behavior.

## Verification

- `cd server && pnpm exec vitest run
src/services/heartbeat/runtime-execution.test.ts
src/services/native-runtime/remote-warm-session-recovery.test.ts`: 57
passed.
- `npx tsc --noEmit -p .` in `server` reports no errors.
- New tests cover the log line at the throw site, and the helper's
output for attach failures and other failures.
- **Not covered:** the new event in `heartbeat.ts` has no test of its
own. It uses the tested helper.
- **Not verified:** I have not confirmed that agents read this log when
they diagnose a failure.
- **Not run locally:** the full test suite. CI provides that coverage.

## Risks

- **Failure text in the log:** when the failure has no structured
message, the log line includes the whole failure string. It passes
through the run log's identity redaction. The event uses the stored
failure that the executor already redacted.
- **Link stability:** the link depends on the docs heading. Renaming
that heading would break the anchor but not the page link.
- **Migrations:** this change has no migration.

> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be reworked.

## Model Used

- Anthropic Claude Opus 5.5 (`claude-opus-5-5`), through Claude Code,
with tool use and code execution. A human directed the design and
approved the scope.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-04 14:36:17 -08:00

190 lines
13 KiB
TypeScript

import { and, eq } from "drizzle-orm";
import { agents, companies, environmentLeases, heartbeatRuns, issues, nativeRunFinalizations, type Db } from "@paperclipai/db";
import type { AdapterExecutionTarget } from "@paperclipai/adapter-utils/execution-target";
import { parseNativeExecutionInput, type NativeExecutionInput, type PersistedNativeSession } from "../../vendor/paperclip-runner/index.js";
import { currentNativeControllerIdentity, evaluateNativeControllerTakeover } from "./native-restart-recovery.js";
import { NativeControllerRecoveryWaitingError, readNativeControllerRecovery } from "./native-controller-recovery.js";
export type RemoteWarmSessionRecovery = {
identity: Record<string, string>;
checkpoint: PersistedNativeSession;
previousExecution: NativeExecutionInput;
reuseProcess: boolean;
remote: { providerLeaseId: string; remoteCwd: string };
revalidate: (leaseOwner: string) => Promise<void>;
};
/** Protocol events that mean work started and may still be live or waiting. */
const POST_TERMINAL_LIVE_WORK = new Set(["turn.started", "turn.accepted", "run.attached", "tool.execution.started",
"semantic_tool.input", "runtime_request.created"]);
const record = (value: unknown): Record<string, unknown> =>
value !== null && typeof value === "object" && !Array.isArray(value) ? value as Record<string, unknown> : {};
function postTerminalEvents(control: Record<string, unknown>): unknown[] {
const events = Array.isArray(control.committedEvents) ? control.committedEvents : [];
const terminalIndex = events.findLastIndex(event => record(event).eventType === "run.terminal");
return terminalIndex < 0 ? [] : events.slice(terminalIndex + 1);
}
/** A session is settled when its commands finished and nothing is live or waiting.
* Events after the terminal describe the past, so refusing cannot undo them: they
* are reported instead. Every event must still authenticate as this session's. */
export function isSettledRemoteWarmControlState(control: Record<string, unknown>, identity: Record<string, unknown>): boolean {
const commands = control.commands;
const events = control.committedEvents;
if (!Array.isArray(commands) || commands.some(command => record(command).status !== "completed") ||
!Array.isArray(events) || Object.entries(identity).some(([key, value]) => record(control.identity)[key] !== value)) return false;
const terminalIndex = events.findLastIndex(event => record(event).eventType === "run.terminal");
if (terminalIndex < 0) return false;
const matchesSession = (event: unknown) => {
const envelope = record(record(event).envelope);
return envelope.runId === identity.runId && envelope.normalizedSessionId === identity.normalizedSessionId;
};
if (!matchesSession(events[terminalIndex])) return false;
return events.slice(terminalIndex + 1).every(event => {
const envelope = record(record(event).envelope);
const notification = record(envelope.payload);
const payload = record(notification.payload);
const eventType = record(event).eventType;
// An active goal keeps a run going, and its next turn can start at any moment.
const live = (typeof eventType === "string" && POST_TERMINAL_LIVE_WORK.has(eventType)) ||
((eventType === "session.goal.snapshot" || eventType === "session.goal.updated") &&
(payload.workingNow !== false || record(payload.goal).status === "active"));
return !live && matchesSession(event) &&
Object.entries(identity).every(([key, value]) => envelope[key] === value) &&
(notification.schema === "paperclip.prp.event.v1" || notification.schema === "paperclip.prp.event.v2") &&
notification.eventType === eventType &&
notification.sourceInstanceId === identity.runnerInstanceId && notification.sourceKind === "runner" &&
notification.runId === identity.runId && notification.normalizedSessionId === identity.normalizedSessionId;
});
}
/** Event types recorded after the prior run's terminal, for the next run's log. */
export function summarizePostTerminalEvents(control: Record<string, unknown>): string | null {
const counts = new Map<string, number>();
for (const event of postTerminalEvents(control)) {
const type = String(record(event).eventType);
counts.set(type, (counts.get(type) ?? 0) + 1);
}
return counts.size === 0 ? null : [...counts].map(([type, count]) => `${type} x${count}`).join(", ");
}
/** A failed attachment suspends the Runner under the attaching run while the journal
* still names its predecessor. A same-run retry cannot prove either owner, so report
* the failure that caused this state rather than the resulting identity mismatch. */
export function describeOwnFailedAttachment(recoveryHistory: unknown): string {
const history = Array.isArray(recoveryHistory) ? recoveryHistory : [];
const cause = record(history.at(-1)).stderrTail;
return `runner_state_identity_mismatch: prior_attach_failed: ${typeof cause === "string" && cause ? cause : "unknown"}`;
}
const SETTLEMENT_DOC = "https://github.com/paperclipai/paperclip/blob/master/doc/execution-semantics.md#session-settlement-after-a-run";
/** Points whoever diagnoses a failed attachment at reviewed docs. The stored
* failure is already redacted and bounded; other failures produce no pointer. */
export function describeFailedAttachment(failureMessage: unknown): string | null {
if (typeof failureMessage !== "string" || !/\brun\.attach (failed|requires)\b/.test(failureMessage)) return null;
const encoded = /"message":("(?:[^"\\]|\\.)*")/.exec(failureMessage)?.[1];
let cause = failureMessage;
try { if (encoded) cause = JSON.parse(encoded) as string; } catch { /* Keep the whole failure. */ }
return `This run could not attach to its provider session: ${cause}. A later run may start a new provider thread. See ${SETTLEMENT_DOC}`;
}
export async function inspectRemoteWarmSessionRecovery(input: {
db: Db;
execution: NativeExecutionInput;
target: Extract<AdapterExecutionTarget, { kind: "remote" }>;
identity: Record<string, unknown>;
readControlState: () => Record<string, unknown>;
matchesExecution: (previous: NativeExecutionInput) => boolean;
loadCheckpoint: (previous: NativeExecutionInput) => PersistedNativeSession | null;
canReuseProcess: (previous: NativeExecutionInput) => boolean;
}): Promise<RemoteWarmSessionRecovery> {
const deny = (): never => { throw new Error("runner_state_identity_mismatch: prior_owner_terminal_state_indeterminate"); };
const { binding } = input.execution;
const { target } = input;
const identity: Record<string, string> = {};
for (const key of ["runId", "normalizedSessionId", "runnerInstanceId", "environmentLeaseId", "turnId", "itemId"]) {
const value = input.identity[key];
if (typeof value !== "string" || !value) return deny();
identity[key] = value;
}
if (target.transport !== "sandbox" || !target.leaseId || !target.environmentId ||
!target.providerKey || !target.sandboxLeaseAcquisition || target.sandboxLeaseAcquisition.outcome !== "resumed" ||
identity.runId === binding.runId) return deny();
const providerLeaseId = target.sandboxLeaseAcquisition.providerLeaseId;
const [prior] = await input.db.select({ run: heartbeatRuns, owner: nativeRunFinalizations })
.from(heartbeatRuns).innerJoin(nativeRunFinalizations, eq(nativeRunFinalizations.runId, heartbeatRuns.id))
.where(and(eq(heartbeatRuns.id, identity.runId), eq(heartbeatRuns.companyId, binding.companyId),
eq(heartbeatRuns.agentId, binding.agentId), eq(heartbeatRuns.nativeIssueId, binding.issueId)));
if (!prior || prior.run.status !== "succeeded" || !prior.owner.resultId || !prior.owner.controllerPid ||
!prior.owner.controllerProcessStartedAt || !prior.owner.controllerBootId) return deny();
const previous = parseNativeExecutionInput(prior.run.runnerProfileJson?.nativeExecutionInput);
if (previous.binding.runId !== identity.runId || !input.matchesExecution(previous)) return deny();
if (prior.owner.leaseOwner && prior.owner.leaseExpiresAt && prior.owner.leaseExpiresAt > new Date()) {
throw new NativeControllerRecoveryWaitingError(prior.owner.leaseExpiresAt);
}
if (prior.owner.phase !== "committed" || prior.owner.leaseOwner) return deny();
const checkpoint = input.loadCheckpoint(previous);
if (!checkpoint || checkpoint.identity.runId !== identity.runId ||
checkpoint.identity.companyId !== binding.companyId || checkpoint.identity.issueId !== binding.issueId ||
checkpoint.identity.agentId !== binding.agentId || checkpoint.identity.sessionId !== identity.normalizedSessionId ||
checkpoint.driverKind !== previous.session.driverKind || !checkpoint.providerSessionId ||
checkpoint.activeTurnId !== null || checkpoint.pendingRuntimeRequests?.length !== 0) return deny();
const control = input.readControlState();
if (!isSettledRemoteWarmControlState(control, identity)) return deny();
const assertPriorStopped = async () => {
const persistedRecovery = readNativeControllerRecovery(prior.run.runnerProfileJson?.nativeControllerRecovery,
prior.run.runnerProfileJson?.sessionCheckpoint);
// A PID in another webserver container cannot prove life or death. New
// durable controllers are fenced by the database lease and current turn.
const stopped = await evaluateNativeControllerTakeover({ owner: prior.owner, now: new Date(), requireStoppedController: true,
remoteLeaseFenced: persistedRecovery !== null });
if (!stopped.allowed) return deny();
};
await assertPriorStopped();
const priorLeases = await input.db.select().from(environmentLeases).where(and(
eq(environmentLeases.companyId, binding.companyId), eq(environmentLeases.heartbeatRunId, identity.runId),
eq(environmentLeases.issueId, binding.issueId), eq(environmentLeases.environmentId, target.environmentId),
eq(environmentLeases.provider, target.providerKey), eq(environmentLeases.providerLeaseId, providerLeaseId),
));
if (priorLeases.length !== 1 || priorLeases[0].metadata?.remoteCwd !== target.remoteCwd ||
!["released", "expired", "active"].includes(priorLeases[0].status)) return deny();
return {
identity, checkpoint, previousExecution: previous, reuseProcess: input.canReuseProcess(previous),
remote: { providerLeaseId, remoteCwd: target.remoteCwd },
revalidate: async leaseOwner => {
await assertPriorStopped();
const controller = await currentNativeControllerIdentity();
const [current] = await input.db.select({ run: heartbeatRuns, owner: nativeRunFinalizations, issue: issues, agent: agents, company: companies })
.from(heartbeatRuns).innerJoin(nativeRunFinalizations, eq(nativeRunFinalizations.runId, heartbeatRuns.id))
.innerJoin(issues, eq(issues.id, heartbeatRuns.nativeIssueId))
.innerJoin(agents, and(eq(agents.id, heartbeatRuns.agentId), eq(agents.companyId, heartbeatRuns.companyId)))
.innerJoin(companies, eq(companies.id, heartbeatRuns.companyId))
.where(and(eq(heartbeatRuns.id, binding.runId), eq(heartbeatRuns.companyId, binding.companyId),
eq(heartbeatRuns.agentId, binding.agentId), eq(heartbeatRuns.nativeIssueId, binding.issueId)));
const [lease] = await input.db.select().from(environmentLeases).where(and(
eq(environmentLeases.id, target.leaseId!), eq(environmentLeases.companyId, binding.companyId),
eq(environmentLeases.heartbeatRunId, binding.runId), eq(environmentLeases.issueId, binding.issueId),
));
const [priorOwner] = await input.db.select().from(nativeRunFinalizations).where(and(
eq(nativeRunFinalizations.runId, identity.runId), eq(nativeRunFinalizations.companyId, binding.companyId),
eq(nativeRunFinalizations.issueId, binding.issueId),
));
if (!current || current.run.status !== "running" || current.run.resultJson?.nativeCancellation ||
current.company.status !== "active" || current.agent.lifecycleState !== "ready" ||
["paused", "terminated", "pending_approval"].includes(current.agent.status) ||
!priorOwner || priorOwner.leaseOwner || priorOwner.phase !== "committed" ||
priorOwner.resultId !== prior.owner.resultId || priorOwner.controllerBootId !== prior.owner.controllerBootId ||
current.owner.leaseOwner !== leaseOwner || !current.owner.leaseExpiresAt ||
current.owner.leaseExpiresAt <= new Date() || current.owner.resultId ||
current.owner.controllerBootId !== controller.bootId || current.owner.controllerPid !== controller.pid ||
current.issue.executionRunId !== binding.runId || current.issue.assigneeAgentId !== binding.agentId ||
current.issue.status !== "in_progress" || !lease || lease.status !== "active" ||
lease.environmentId !== target.environmentId || lease.provider !== target.providerKey ||
lease.providerLeaseId !== providerLeaseId || lease.metadata?.remoteCwd !== target.remoteCwd) return deny();
},
};
}