Drop interrupted workers from live agents

[TheGreatAxios]

Summary

• Drops interrupted workers from the live Agents strip after the same linger window as done, failed, and cancelled
• Stops sticky clocks and fleet-running counts from treating an interrupted worker as live while leftover tools may still be running
• Keeps cancelled linger painted as cancelled, not interrupted

Verification

• bun test src/tui/chrome-state.test.ts passes, including cancelled-linger and interrupted linger/drop cases
• bun run check passes

Fixes CL-7175

[linear-code]

CL-7175

[TheGreatAxios]

lens: critic
verdict: clean

This branch drops interrupted workers from the live agents list after the linger window and keeps leftover tools from pinning a running slot. The cancelled-linger wording regression found on the first pass is fixed. I found no remaining blocking or should-fix defects.

[TheGreatAxios]

Review · Request changes

Stamps finishedAt on interrupt and adds agentLaneIsLive() so the agents strip, sticky poll, and fleet roll-up treat status:"running" + lifecycleStatus:"interrupted" as a lingering row that drops after 4s.

Findings

• src/subagent/session-store.ts:1021 / :1030-1057 + src/tui/agent-progress.ts:132 — interruptOne stamps finishedAt and leaves lifecycleStatus:"interrupted", but followupOne (and the sendInputOne interrupt path at :972-1000) runs the new worker turn without flipping lifecycle back to "running" or clearing finishedAt (only resumeOne at :1076 does, gated to "completed"). During the follow-up turn the session is running/interrupted/finishedAt=<interrupt time>, so once now - finishedAt >= 4000: formatAgentsPanel → null, agentsChromeNeedsSticky → false, fleetProgress.running → 0, while tool events are still arriving. Reproduced (interruptOne → pending followupOne → appendEvent(tool_call.start) at t+9s): branch prints panel: null, sticky: false, fleet.running: 0; main shows the lane ● explorer d · interrupted · run_shell still running. This is the interrupt_agent → followup_task / send_input {interrupt:true} flow exposed by lifecycle-tools.ts:261,319,392. On followup start, mutate lifecycle to "running" and drop finishedAt the way resumeOne does, then let the existing .then re-stamp on completion (check isSoftInterrupted in agent-fleet.ts:799-811 first).
• src/subagent/agent-fleet.test.ts:925-968 — the new test passes with session-store.ts, agent-fleet.ts, and lifecycle-tools.ts checked out from origin/main (bun test src/subagent/agent-fleet.test.ts -t "after interrupt_agent wait-status" → 1 pass), because list_agents status comes from fleetRecords, which interrupt_agent already flipped. It does not exercise this change.

Notes

• s.finishedAt ?? now() keeps the second interruptOne from agent-fleet.ts:741 idempotent — correct.
• Checks: bun run typecheck, bun run lint, bun test src/tui/chrome-state.test.ts src/tui/agent-progress.test.ts src/subagent/session-store.test.ts src/subagent/agent-fleet.test.ts src/subagent/lifecycle-tools.test.ts src/subagent/followup-live-agent.test.ts src/tui/shell.test.ts (169 pass).