tcgflow-session-reportlisted

# `/tcgflow-session-report` — author a session post-mortem The Cockpit's **Session report** page renders the *charts* deterministically (token classes, cost waterfall, tool-calls-by-type, per-turn trace) and its "Open report ↗" exports them as standalone HTML. This command produces the **editorial layer the charts can't**: a narrative headline, a "what happened" summary, and **ranked optimization recommendations with $ savings** — the prose that needs a model, not a server. ## Inputs A task ID (e.g. `ES-6965`). Its orchestrated **Runs** live at `.tcgstackflow/runs/{TASK-ID}/{run-id}.md`; each run-record's frontmatter carries the `session_id`. The rich source is the **Claude Code session JSONL** at `~/.claude/projects/*/{session_id}.jsonl` (per-turn `usage`, `tool_use` events, timestamps, model). ## Procedure 1. **Locate the runs.** Read `.tcgstackflow/runs/{TASK-ID}/*.md`; collect each run's `session_id` and `role`. If there are none, stop and tell the user the task has no orchestrated runs to report on. 2. **Find + parse each session JSONL.** `find ~/.claude/projects -name "{session_id}.jsonl"`. Per file, aggregate across `assistant` records: token classes (`input`, `output`, `cache_read_input_tokens`, `cache_creation_input_tokens`), `tool_use` counts by name, timestamps (wall-clock), the model. **If a JSONL is missing on this machine, say so — do not fabricate per-turn data** (fall back to the run-record frontmatter totals and note the trace is unavailable). [[no-fabri