feat(agent): long-running task loop with LLM-planned proposals

[robbie-c]

Summary

Adds an auto-continuing turn loop the agent can enter for tasks with an objective, measurable success criterion (test pass count, bundle size, error count, etc.). Inspired by Anthropic's Ralph Wiggum plugin but redesigned around the desktop agent framework instead of a bash loop.

How it works

1. User types /long-running-task <description> (or the agent self-initiates).
2. A brief planning prompt goes to the agent, with planning instructions attached as hidden _meta context. The agent explores first — reads relevant files, runs measurement commands, asks clarifying questions via AskUserQuestion only when intent is genuinely ambiguous.
3. The agent emits a proposal as a <long-running-task-config>{...}</long-running-task-config> JSON block (goal, successCriterion, marker, maxIterations, approach).
4. The renderer parses the block on end_turn and shows an inline confirmation card above the prompt input — editable fields, Start / Edit / Cancel.
5. On Start, the renderer calls a new tRPC mutation that sets server-side state, then sends a kickoff prompt.
6. After each end_turn while the task is active, the loop in claude-agent.ts pushes a continuation user message into session.input and keeps the while-loop alive instead of returning. The agent exits when:
   • It emits the configured marker (default <TASK_COMPLETE>) in its assistant text
   • It hits the max-iteration cap (default 20) — gets one wrap-up turn before exiting
   • The user clicks Stop loop in the pill
   • The user cancels normally
7. Mid-loop user messages pause auto-continuation for that turn so the user can steer.

A pill above the prompt input shows iteration count + goal + Stop button while a task is active.

Files of note

• packages/agent/src/adapters/claude/long-running-task/utils.ts — schemas, continuation builder, proposal parser, broadcast helpers
• packages/agent/src/adapters/claude/claude-agent.ts — loop hook in the result.subtype === "success" branch + new START_LONG_RUNNING_TASK / STOP_LONG_RUNNING_TASK extMethods
• packages/agent/src/adapters/claude/session/instructions.ts — appended system prompt teaching the proposal workflow
• apps/code/src/renderer/components/long-running-task/ — pill + proposal card
• apps/code/src/renderer/stores/longRunningTaskStore.ts — Zustand store keyed by taskRunId

Known limitations

• Local sessions only. The slash command early-returns with a toast on cloud sessions — extMethod doesn't work over the cloud path.
• Sentinel marker is trust-based. The agent could in theory emit the marker without actually verifying. The system prompt addendum is explicit about this; the iteration cap is the hard safety net. We can revisit with structural verification later if the trust-based version misfires.

Test plan

• /long-running-task with an iterative task (e.g. fix N failing tests). Agent explores, asks if needed, emits proposal block, card renders.
• Approve the proposal card → pill appears with 0/20, loop kicks in.
• Counter increments on each end_turn until marker is emitted or cap is hit.
• Edit the proposal in the card before approving — edited values persist when starting.
• Click Stop loop mid-flight — current turn finishes, no more continuations injected.
• Send a steering message mid-loop — agent responds to user, loop stays active and continues after.
• Hit the iteration cap deliberately (set max=2 in a clearly-impossible task) — wrap-up turn runs.
• /long-running-task on a cloud task → toast says not supported.
• Cancel the entire prompt mid-loop → loop exits cleanly.

---

Created with PostHog Code