feat(core): the agent indicator in the diagnostics log (#152)

[Maxaubert]

Closes #152.

The diagnostics log (#140) now records the agent indicator, so a wrong mark like #151 (restored tabs showed Finished after a restart) can be read from diag.jsonl rather than guessed. Quiet level, always on.

What changed

• core/renderer/lib/agentDiag.ts (new): the record, pure helpers plus the page's one instance.
  • agent-hook: every OSC 777 prism-agent signal (id, state, kind). A run of the same state on a tab is folded: the first is written at once, the rest counted and written as one line with repeats when the state changes, the tab closes, or after 5 s with no signal from it (t is the last repeat's). A turn of 40 tool calls is two lines.
  • agent-title: the title's meaning (idle / working / starting / question / none), only when it changes, never the text and never each spinner frame.
  • agent-mark: each move of a tab's mark (from, to, held, why, agent, restored ms when within a minute of a restore). why is the rule that fired: hook <state> (, box on screen, , tab in front), spinner after <phase>, idle title after work (an Esc), title <state>, poll found/lost agent, screen read: question box, answer key, box gone, output scored working / output went quiet, stopped working while away, tab looked at, working again, agent gone, closed.
  • agent-restore: a tab restored with an agent resume (id, cwd, resume: true).
• useAgentIndicator.ts: noteWhy at each decision point and one read-only effect that writes the moves. No rule changed (the gate e2e below all pass unchanged).
• diag.ts: record(k, fields) for a page's own record kinds; diagIpc.ts: main accepts the four agent-* kinds from the page (and no other agent- kind).
• termActivity.markResume(id, session, cwd?): logs agent-restore; App passes the folder.
• tools/diag.mjs: --agent (the timeline: agent lines, tab and shell crumbs, sessions, marks), --tab <id>; --all and --kinds show them; the default problem view shows agent-mark / agent-restore among the context before a problem. --help prints the whole header.
• docs/diagnostics.md (kinds table, a section on reading the marks), PRIVACY.md (one clause: the indicator's states, never screen, title or conversation text).
• Versions: core 0.28.1 -> 0.29.0, app 0.35.1 -> 0.35.2.

Prism gets this too

This is a core/ change, so Prism gets the record when it takes this core. Nothing to wire beyond what #140 already needs (startDiag in the page): the hook and title lines and the marks come from useAgentIndicator. Optional: pass the tab's folder as the third argument of markResume so agent-restore carries cwd (without it, cwd is null).

Tests

• Unit (test-first): agentDiag.test.ts (folding, the quiet sweep, kinds, the meaning-change filter, mark order and held marks, why freshness, restore, close) and a pageLine case for the new kinds.
• e2e agentDiag (new): a shell stands in for Claude; reads the scenario profile's logs\diag.jsonl and asserts the hook sequence ["working","workingx2","done","question","working","done"], the mark moves with their why (none>working: hook working, working>done: hook done, done>none: tab looked at, none>question: hook question, question>working: hook working, tab in front, working>none: hook done, tab in front), titles ["idle","working","none"], and that a marker string on the screen and in the title never reaches the file.
• Gates: typecheck, lint (0 errors), npm test 110 files / 1571 tests, e2e agentHooks 37/0, attention 25/0, indicator 12/0, indicatorStyles 74/0, diagLog 26/0, agentDiag 12/0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHHaWKR4M5QtW7Wecyuk4t

[Maxaubert]

Adversarial review, fixed in 1f685f7 (tests first, each failed on the old code):

• Folding did not hold for real turns: a 5 s quiet gap ended a run, and the model's thinking between tool calls is usually longer, so nearly every hook was its own line. A run now folds across gaps, writes its count at most every 30 s (and at once on a state change or close), and starts a new line after a minute of silence. Ten minutes of tool calls every 8 s: 150 signals, at most 23 lines.
• kind was pty bytes (any [a-z0-9_]{1,40} passed): only Claude Code's own StopFailure kinds are written, anything else is other.
• Codex title flapping: its child processes set titles between its spinners, which wrote a none/working pair each time. none is written only once it has held 2 s.
• Flood cap: any program can print the OSC or a title; at most 40 hook/title/mark lines per tab per 10 s, the next line carries dropped (diag.mjs shows it).
• Stale why: each new note refreshed the whole list's time, so a busy tab kept older rules in the reason forever. Only rules noted in the last second are listed now.

Behaviour: no indicator rule changed (diff of useAgentIndicator.ts is notes and one read-only effect). Volume: 8 working Claude tabs are some tens of lines a minute, a few hundred KB an hour of the 10 MB.

Gates: typecheck, lint (0 errors), npm test 1577 pass; e2e one at a time: agentDiag 12, agentHooks 37, attention 25, indicator 12 + indicatorStyles 74, all pass. CI green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FHHaWKR4M5QtW7Wecyuk4t