Add a task time-profiler: llm/tool/operator-wait gap analysis over transcripts

[dkrattiger]

Summary

Answers the operator goal ("profile tasks so I know how long the agent is spending working on various things including waiting on tests, waiting on other calls, vs waiting on the llm to respond") retroactively, with no agent-side instrumentation — pure gap analysis over the timestamps already in every claude session transcript. Plan: plan.md on this task.

• profiler/ (new, pure/LLM-free package): categories.py is the tool-time category table — one ordered list of matchers (Bash sub-classified by regex into tests/vcs/deps/pty-verify; Read/Write/Edit/Grep/Glob → code-nav; Task/Agent → subagents — one bucket, its sidechain is never walked separately; mcp__* → orchestration; the rest → other-tools) — easy to extend in one place. parse.py's profile_transcripts(paths) -> dict merges same-message.id assistant lines into one turn (claude streams thinking/text/tool_use as separate JSONL lines per block), then classifies every gap: user → assistant is always LLM time; assistant → user is tool time (when the turn has tool_use) or operator wait (when it doesn't — the agent stopped and handed back to the human). AskUserQuestion spans are routed to operator wait, not a tool category, since the turn-flip hooks make that a blocked-on-human wait, not tool execution. A same-session restart (two consecutive user-type lines with no assistant turn between them — e.g. claude's "You were interrupted. Continue." landing right after an abandoned tool_result) is billed to operator wait too, rather than silently vanishing. Multiple session files (restarts/phases) are concatenated by timestamp, with between-session gaps reported as their own bucket. Everything is defensive — malformed/old-format lines are skipped, never crash the parser; unattributed time is surfaced, not hidden.

• sessionservice/transcripts.py: the transcripts live in a per-task named Docker volume (panopticon-config-<task_id>) that outlives the container — there was no existing host-accessible path to it. task_session_paths reads it via docker run (find/cat, reusing the already-built panopticon-base image), behind an injectable command-runner. It checks the volume exists first — a bare docker run --volume <missing-name> would otherwise silently create an empty one. Local-runner scope only for v1 (remote hosts are a natural follow-up via the same ssh-wrap pattern terminal/attach.py::attach_command already uses).

• CLI: panopticon profile <task-id-or-slug> — total wall / operator wait / between-sessions / agent-active split by category (counts + totals + top 5 longest individual tool calls). --all-tasks aggregates per-repo totals and medians, e.g. answering "what fraction of tarot agent time is pytest?" in one command.

• Dashboard: a new P hotkey time-profiles the highlighted task on demand and shows a one-line summary in the detail pane (agent 4.3h: llm 43% tests 19% tools 38% | waited on user 9.2h). Deliberately not automatic (unlike the rest of the detail pane) — it shells out to docker per task, which is too expensive to redo on every arrow-key highlight.

Known limitation / pending

This container has no docker access, so I couldn't run the profiler against real tarot task transcripts myself. Per discussion on the task, I'll add real-data output from 3 completed tarot tasks to this description before merge.