Active tools, background work and stopped sessions

The shared TUI, web and desktop source separates native execution waits from questions, permission requests, provider errors and a ready input prompt. It uses the existing Session.lifecycleState, lifecycleActivity and native goal observations; there is no second session state machine or model-based classifier.

Display Evidence Meaning
⚙️ RUNNING TOOL · Bash The latest observed root Claude/Codex phase is PreToolUse The agent is working through a tool call. A pre-call hook can precede permission checks; it does not prove that an OS process has started or that the agent is waiting for external work.
⏳ WAITING FOR BACKGROUND WORK Antigravity Stop with fullyIdle: false, without an error stop The native CLI reports remaining work. Its process and child identity are not reported.
WORKING · LAST TOOL FAILED · Bash Root PostToolUseFailure One call failed, but the turn is still working and can recover.
⚠️ CLI RUNNING · TURN FAILED Native failed lifecycle while the CLI process remains running The provider ended the turn with an error, but its terminal remains available. This does not mean that the process exited. A subsequent native working/completed event clears the failed-turn status. Existing model-capacity recovery can supply a more specific status.
⛔ STOPPED · GOAL BLOCKED A fresh blocked native goal for this exact account, conversation and process generation, at a stopped prompt The CLI explicitly reports that its goal is blocked. It does not override new work, questions or permission requests.
💬 QUESTION · NEEDS YOUR ANSWER Explicit native user-input tool A user response is required.
✋ NEEDS APPROVAL Native permission request A permission decision is required.
💬 READY / NEEDS INPUT Ordinary completed native turn The prompt is ready. This alone does not prove a problem or successful task completion.

Active tools keep the working gear icon. Explicit background waits use cyan; questions and permissions use yellow; provider errors use red. These are fixed status labels, not an endless progress animation. The same renderer and classifier are used by TUI, web and Tauri source.

Freshness and limitations

A post-call event clears the displayed pre-call phase. This is the latest native hook phase, not a scheduler or an inventory of every parallel invocation: another call's result can restore the generic working label while an earlier call is still outstanding. A background agent launch that has already returned is not treated as an outstanding tool call. Merely seeing an active subagent, Docker process, listening service, CPU usage or a quiet terminal never establishes that the parent is waiting for it. Tool names are shown, but arguments, command lines, output, exception text and credentials are discarded.

Native pre-call hooks do not expose the full process tree. After five minutes without another native wait event, a still-observed lifecycle becomes RUNNING · WAIT NOT RECONFIRMED, with a question mark. If host/lifecycle proof itself becomes stale, the existing unknown-activity status takes precedence. Neither case is declared blocked, idle or actively executing a known process. The execution host may refresh collector availability independently; that does not reset this native wait-event age.

A plain assistant sentence such as “I am blocked” cannot create a confirmed blocked status. If the CLI exposes neither a native error, question/permission nor a current blocked goal, devctrl retains its ordinary ready/unknown status. There are no new notifications or attention categories for active tool calls.

Existing managed Claude/Codex hook files are refreshed atomically by the host observer without sending input or restarting the CLI. The next actual native event supplies the new activity. Older tool:<name> observations remain compatible but do not retroactively establish a wait. External package reporters receive this change during their normal integration preparation/update; sessions still using an older Antigravity reporter retain their previous generic working status until that reporter is updated. The CLI reads the script on each hook; no task is injected to force a new event.

API and MCP

Existing session inventory carries the same fields, so authorized MCP/operator clients can inspect the phase without another endpoint. New bounded activities are awaiting-tool[:<name>], tool-failed[:<name>], turn-failed, and background-work; tool[:<name>] continues to mean a completed call. The legacy awaiting-tool activity maps to running_tool in the pure sessionExecutionObservation client helper, which checks source, freshness and goal identity before interpreting them. Public lifecycle states and authorization semantics do not change.

Official event contracts: