Agent indicator: permission prompts often miss the Question line, and no Working mark after approving #148

Reported by the owner on 2026-10-10 from Prism Terminal Stable 0.34.0, which already has the hook signals from #131. It came with a screenshot of the Downloads tab.

What happens

1. The question line is often missing. When Claude Code asks for permission (for example before a gh pr merge), the tab often does not get the blue question line.
2. The working mark is missing after approval. After the permission was accepted, Claude went on working: the screenshot shows Moseying… (7m 12s) and a shell command running for 33 s. The tab still showed no working indicator, and it stayed that way while the work continued.

Expected

• A permission prompt raises Question every time.
• Accepting it goes straight back to Working, until Stop.

Leads to check (not verified)

• No hook between approval and the tool's end. After an approval no hook fires until PostToolUse, which comes only when the tool finishes. So a long shell command leaves the tab with no Working mark for as long as it runs. agentHookState should treat the approval as work again. The key that answers the prompt is the only signal (Question mark should stay until the question is answered #144's answersQuestion), so after that key it should go to working, not idle.
• The prompt was looked at, which cleared it on 0.34.0. On 0.34.0 the question cleared once the tab was viewed. fix(core): a question mark stays until the question is answered (#144) #145 (merged in 0.34.1 and later) keeps it until it is answered, but it may still drop to "nothing" instead of "working".
• Auto mode's permission prompts may not go through PermissionRequest or Notification permission_prompt. The session in the screenshot runs in auto mode, where the classifier is the one that asks. Measure which hooks fire for one of these prompts.
• The plugin may not have been loaded in that tab. It could be a tab restored or a warm shell adopted before the plugin was handed over. Check CLAUDE_CODE_PLUGIN_DIRS in that shell.

Done when

• An e2e (or a measured pty run) shows: a permission prompt gives Question; approving gives Working at once; Working holds through a long tool call; Stop gives Finished.
• The same holds for an auto-mode classifier prompt.

Related: #131 (hooks), #144 / #145 (question stays until answered), #143 (indicator styles).