fix: retry sends after stale agent stop

[yaukwan]

No description provided.

[greptile-apps]

Greptile Summary

This PR adds automatic retry logic to two send_agent_message code paths so that a message is re-dispatched when the underlying provider session becomes stale after an agent stop. A new waitForAgentRunStartOrThrow helper inlines the existing waitForAgentRunStartWithTimeout utility, removing the need for that import.

• session.ts: Extracts dispatchAgentMessagePrompt, retrySendAgentMessageAfterSessionReload, and waitForAgentRunStartOrThrow helpers, then wires them into both the RPC (handleSendAgentMessage) and WebSocket request (handleSendAgentMessageRequest) paths with a single-shot reload-and-retry on failure.
• agent-manager.test.ts: Adds a focused unit test using a FailsOnceBeforeTurnSession stub that verifies reloadAgentSession recovers the stale state and the next runAgent succeeds.
• daemon-client.e2e.test.ts: Removes two unused imports left over from a previous refactor.

Confidence Score: 4/5

Safe to merge — the retry path is single-shot with proper error surfacing on failure, and the new test covers the core recovery flow end-to-end.

The retry guard fires for any active agent regardless of error type, so non-stale-session failures also trigger a session reload and run cancellation. The effect is bounded but the guard could be more precise. There is also a dead text parameter in two helper method signatures.

packages/server/src/server/session.ts — specifically the shouldRetrySendAfterStartFailure condition and the unused text fields in the two new helper input types.

Important Files Changed

Filename	Overview
packages/server/src/server/session.ts	Adds retry logic (with session reload) for two send-message paths; the guard condition is broader than the targeted scenario and two helper methods carry an unused text parameter.
packages/server/src/server/agent/agent-manager.test.ts	Adds a well-structured test that verifies reloadAgentSession recovers a stale session and that a subsequent runAgent succeeds; no issues found.
packages/server/src/server/daemon-client.e2e.test.ts	Removes two unused imports (AgentSession, AgentStreamEvent) — clean housekeeping with no functional impact.

Sequence Diagram

sequenceDiagram
    participant Client
    participant Session
    participant AgentManager
    participant Provider

    Client->>Session: send_agent_message
    Session->>AgentManager: sendPromptToAgent / waitForAgentRunStart

    alt Success
        AgentManager->>Provider: startTurn()
        Provider-->>AgentManager: "{ turnId }"
        AgentManager-->>Session: run started
        Session-->>Client: "{ ok: true }"
    else Failure (stale session)
        Provider-->>AgentManager: throws session unavailable
        AgentManager-->>Session: error
        Session->>Session: shouldRetrySendAfterStartFailure?
        note right of Session: checks agent.session != null
        Session->>AgentManager: reloadAgentSession()
        AgentManager->>AgentManager: cancel in-flight run
        AgentManager->>Provider: resumeSession / createSession
        Provider-->>AgentManager: new session
        Session->>AgentManager: sendPromptToAgent (retry)
        AgentManager->>Provider: startTurn()
        Provider-->>AgentManager: "{ turnId }"
        AgentManager-->>Session: run started
        Session-->>Client: "{ ok: true }"
    else Retry also fails
        AgentManager-->>Session: retryError
        Session->>Session: handleAgentRunError
        Session-->>Client: "{ ok: false, error }"
    end

Loading

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into fix/send-after-..." | Re-trigger Greptile}

[greptile-apps]

Retry condition is unconditionally true for all active agents

agent.session != null holds for virtually every active agent, so this guard returns true for any error thrown by sendPromptToAgent — including storage failures, lifecycle state violations, and transient errors that have nothing to do with a stale provider session. When one of those unrelated errors triggers the condition, reloadAgentSession is still called, which cancels any in-flight run and tears down the current session before re-dispatching. A more targeted check (e.g. matching a specific error type/message emitted when the provider session becomes unavailable, or inspecting a dedicated lifecycle field) would confine the reload to the scenario the PR actually intends to fix.

[greptile-apps]

The text field is declared in the input shapes for both helpers but is never referenced inside their bodies — only prompt, agentId, messageId, and runOptions are forwarded to sendPromptToAgent. This is dead code at the parameter level and conflicts with the project's noUnusedLocals / noUnusedParameters discipline.

Suggested change

	private async retrySendAgentMessageAfterSessionReload(input: {
	agentId: string;
	text: string;
	prompt: AgentPromptInput;
	messageId?: string;
	runOptions?: AgentRunOptions;
	}): Promise<void> {
	private async retrySendAgentMessageAfterSessionReload(input: {
	agentId: string;
	prompt: AgentPromptInput;
	messageId?: string;
	runOptions?: AgentRunOptions;
	}): Promise<void> {