Add semantic status updates for Codex and OpenCode

[rabanspiegel]

Summary

• add Codex semantic status updates to the shared agent status infrastructure
• add OpenCode semantic status updates and drive working state from real submitted input instead of PTY output parsing
• keep the existing unread/update behavior and shared sidebar/tab indicators for these providers

Testing

• pnpm run type-check
• pnpm exec vitest run src/test/renderer/agentStatusStore.test.ts src/test/main/AgentEventService.test.ts

---

Note

Medium Risk
Changes how agent status transitions to working by driving it from captured submitted input (and extends event mapping) for multiple providers, which can affect sidebar/unread indicators and perceived agent activity. Risk is limited to renderer status/terminal input handling with added unit coverage.

Overview
Adds semantic agent status support for Codex and OpenCode, extending provider adapters so their agent events can drive waiting/complete/error and user submits can set working.

Refactors AgentStatusStore to mark working immediately on submit (removing PTY-output-based confirmation), and updates terminal/injection paths to only call markUserInputSubmitted when a real non-empty submit is detected via new submitCapture tracking. Extracts ANSI stripping into shared stripAnsi and adds unit tests for the new status mappings and submit-capture logic.

^{Written by Cursor Bugbot for commit d673462. This will update automatically on new commits. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
docs	Ready	Preview, Comment	Mar 6, 2026 9:45pm

[greptile-apps]

Greptile Summary

This PR adds semantic status tracking for Codex and OpenCode providers by wiring them into the shared agentStatusStore infrastructure, and replaces the previous PTY-output-parsing approach to determining "working" state with direct tracking of submitted user input across all three supported providers (Claude, Codex, OpenCode).

Key changes:

• providerStatusAdapters.ts — extends the provider allow-lists so that codex and opencode agent events (stop, error, notification) are mapped to the correct AgentStatusKind, and markUserInputSubmitted now returns 'working' for all three providers.
• agentStatusStore.ts — removes the pendingSubmitPtyIds set and the handlePtyData method; markUserInputSubmitted now immediately transitions status to working without waiting for a confirming PTY-output signal.
• TerminalSessionManager.ts — introduces consumeSubmittedInput, a character-level input tracker that determines whether an Enter-press carries actual content, replacing the old PTY-data event hook. The agentStatusStore.handlePtyData call is removed from the PTY data listener. Contains a logic bug in the pending-injection path (see inline comment).
• usePendingInjection.ts — guards markUserInputSubmitted behind text.trim() to avoid false working transitions on empty/whitespace submits.
• useInitialPromptInjection.ts — calls markUserInputSubmitted optimistically just before the PTY input write, correctly covering the initial-prompt path for all supported providers.
• Tests updated to cover the new immediate-working behavior and new provider mappings.

Confidence Score: 3/5

• Safe to merge after fixing the short-circuit logic bug in the pending-injection path of TerminalSessionManager.
• The architectural simplification is clean and well-tested. One concrete logic bug was found in TerminalSessionManager.ts at line 612: the (submittedText || pendingText).trim() expression can silently suppress a markUserInputSubmitted call when submittedText is whitespace-only but pendingText has real content. An additional inherent design limitation exists — terminal history recall (Up Arrow) bypasses currentSubmittedInput tracking because the recalled text arrives as PTY output, not as typed characters — but this is a known trade-off of the input-side approach and not a regression from the previous behavior for Codex/OpenCode (which had no tracking at all before this PR).
• src/renderer/terminal/TerminalSessionManager.ts — specifically the pending-injection condition at line 612.

Important Files Changed

Filename	Overview
src/renderer/lib/agentStatusStore.ts	Removes the pendingSubmitPtyIds set and the handlePtyData method entirely; markUserInputSubmitted now immediately sets status to working rather than waiting for PTY-output confirmation. Clean simplification with no issues.
src/renderer/lib/providerStatusAdapters.ts	Extends provider allow-lists to include codex and opencode in both mapAgentEventToStatus and mapUserInputToStatus. Straightforward, correct change.
src/renderer/terminal/TerminalSessionManager.ts	Introduces consumeSubmittedInput to track typed characters and determine whether an Enter-press carries real content, replacing the prior PTY-output-based confirmation. Contains a logic bug at line 612 where whitespace-only submittedText suppresses a valid pendingText-based markUserInputSubmitted call. Additionally, history-recall via arrow keys is not reflected in currentSubmittedInput (a known design limitation of input-side tracking).
src/renderer/hooks/useInitialPromptInjection.ts	Adds agentStatusStore.markUserInputSubmitted call immediately before the PTY input write; status is optimistically set to working before confirmation. The ordering is safe because a failed ptyInput call leaves sent=false, allowing retry, and PTY exit will eventually reset status to idle.
src/renderer/hooks/usePendingInjection.ts	Adds a text.trim() guard around markUserInputSubmitted so that submitting empty/whitespace text does not incorrectly transition the agent to working. Correct defensive fix.
src/test/renderer/agentStatusStore.test.ts	Updates and extends test coverage for the new immediate working-on-submit behavior for Claude, Codex, and OpenCode; adds notification-mapping tests for all three providers. Removed tests that relied on the now-deleted PTY-data confirmation path. Good coverage.

Sequence Diagram

sequenceDiagram
    participant User
    participant TerminalSessionManager
    participant usePendingInjection
    participant useInitialPromptInjection
    participant agentStatusStore
    participant AgentEventService

    Note over User,AgentEventService: Path A — Direct keyboard input (new consumeSubmittedInput logic)
    User->>TerminalSessionManager: keypress (chars + Enter)
    TerminalSessionManager->>TerminalSessionManager: consumeSubmittedInput() → submittedText
    alt submittedText is non-empty
        TerminalSessionManager->>agentStatusStore: markUserInputSubmitted({ ptyId })
        agentStatusStore-->>agentStatusStore: status → working
    end
    TerminalSessionManager->>ElectronAPI: ptyInput(data)

    Note over User,AgentEventService: Path B — Pending injection (usePendingInjection.sendNow)
    User->>usePendingInjection: sendNow(ptyId, text)
    alt text.trim() non-empty
        usePendingInjection->>agentStatusStore: markUserInputSubmitted({ ptyId })
        agentStatusStore-->>agentStatusStore: status → working
    end
    usePendingInjection->>ElectronAPI: ptyInput(text + "\r")

    Note over User,AgentEventService: Path C — Initial prompt injection
    useInitialPromptInjection->>agentStatusStore: markUserInputSubmitted({ ptyId })
    agentStatusStore-->>agentStatusStore: status → working
    useInitialPromptInjection->>ElectronAPI: ptyInput(prompt + "\n")

    Note over User,AgentEventService: Semantic event updates (Codex / OpenCode / Claude)
    AgentEventService->>agentStatusStore: handleAgentEvent({ type: "notification", notificationType: "idle_prompt" | "permission_prompt" })
    agentStatusStore-->>agentStatusStore: status → waiting (+ unread flag)
    AgentEventService->>agentStatusStore: handleAgentEvent({ type: "stop" })
    agentStatusStore-->>agentStatusStore: status → complete (+ unread flag)
    AgentEventService->>agentStatusStore: handleAgentEvent({ type: "error" })
    agentStatusStore-->>agentStatusStore: status → error (+ unread flag)

Loading

_{Last reviewed commit: 673118a}

[greptile-apps]

Incorrect short-circuit || may suppress markUserInputSubmitted call

(submittedText || pendingText).trim() uses JavaScript's short-circuit ||, which returns submittedText when it is truthy — even if it is whitespace-only (e.g. " "). In that case .trim() yields "" (falsy) and markUserInputSubmitted is never called, even though pendingText is guaranteed truthy and non-empty at this point in the code (the outer if (pendingText && ...) already confirmed it).

Concrete scenario: user presses the spacebar once before an Enter key, triggering the pending-injection path. submittedText would be " " (truthy but whitespace), so " ".trim() is "", the condition is false, and the agent status stays at unknown / its prior state instead of flipping to working.

Suggested change

	if ((submittedText || pendingText).trim()) {
	agentStatusStore.markUserInputSubmitted({ ptyId: this.id });
	}
	if (submittedText?.trim() || pendingText.trim()) {

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}