feat: events table — add interactive_session_start event class with served model + provider

[jakecelentano]

Summary

Today, tms's events table (tms_review.events) captures dispatched-reviewer events with model, provider, session columns, but does NOT capture the interactive session that drove a workstream. So if we want to A/B compare interactive session models (e.g. Claude Code kimi-k3 vs pi/MiniMax-M3 vs future Claude-Code / Max-sub sessions), we have no fleet-side data — only ad-hoc wrap memos and pi footer text.

This is the missing link: a small event class on the events table fires whenever a tmq/aoe session starts, carrying the interactive session's provider + model. Once landed, all the analytics (plan-gate latency, code-review P0-detection, cost-per-task) become sliceable by interactive session model, not just by dispatched reviewers.

Scope

• Add a new event class: interactive_session_start (and interactive_session_end for symmetry).
• Schema change: NONE. Existing columns session, model, provider, event_type, event_timestamp, repo, worktree, payload already cover everything; idempotency UNIQUE on (event_type, aoe_id_prefix, event_timestamp) covers backfill too.
• Hook: at aoe/tmq spawn time (after the worktree is attached, before the agent process boots), emit one interactive_session_start event with the served model+provider from pi's models.json (or equivalent) and the aoe session id as aoe_id_prefix.
• Migration path: introduce the event class in the producer (tms poll, tmq spawn, or pi CLI --record-session-start). Backfill: aoe list can reconstruct historical sessions; models.json on each worktree provides the model. For backfill beyond what the producer captures, scrape existing ~/.local/state/tmq/events.jsonl archives (if any were retained) and the agent-config repo for served-model hints.
• Query surface: add a tms CLI subcommand tms events sessions (or extend tms events transitions) that returns events filtered by event_type='interactive_session_start' and joins on reviewer_runs.model / llm_call_log for the workstream attribution.

Why now

This week's dogfood waves (Jul 19, two wraps) dispatched 14 PRs and recorded 29 reviewer_runs rows, all with panel= attribution. The missing attribution is the interactive session that authored the day. Without it, every "how is model X performing?" question reduces to inference from memos.

Out of scope

• A new "sessions" table — the events table already covers this surface.
• A cost column on the event itself — cost flows through llm_call_log (worktree attribution) and joins from there.
• Per-user attribution (which human kicked off which session) — would need an aoe event for start_session that doesn't exist yet.

Acceptance criteria

• A tmq/aoe spawn emits one interactive_session_start event with session, model, provider, worktree populated
• Event is emitted BEFORE the agent process boots (i.e., the consumer can observe start even if the agent crashes immediately)
• Backfill script (tms backfill session-events or similar) reconstructs the past N days of interactive session starts from aoe list JSON + the worktree's models.json
• New query: tms events sessions --model <id> --since <date> returns the sessions for that model; cross-joins reviewer_runs to attribute panel verdicts to their driving interactive session
• Tests: idempotent insert (ON CONFLICT DO NOTHING on the UNIQUE), backfill idempotency on a re-run, missing-models.json graceful skip

References

• Precedent: distillery's media_digest_runs.model = actually-served (PR distillery#442, memory distillery-provenance-served-model.md) — the same pattern: capture the model that actually served, not the model that should have served.
• Existing migration 006-specialist-composition.sql establishes the precedent for additive column-add migrations on tms_review.
• bogocat/openagent#51 (weekly fleet-digest) is the downstream consumer; this column lands the sliceable dimension.
• Memory: "Fleet runtimes: pi is primary" + "MiniMax-M3 eval — Opus-class; caveat = runtime reliability" make the case for the data without the metric being available.
• Related: tms#101 (verdict capture → reviewer_runs, landed 2026-07-19) did this for dispatched reviewers; this issue extends the same pattern to the interactive session.