web (W11): sessions dashboard — runtime + spend, richer viewer, submission-health panel

[ZIJ]

W11 — Dashboard + Tier-2. Extends the existing Sessions UI (web/) rather than building greenfield. Runtime-agnostic: renders flue and brain-box (claude/codex/pi) sessions alike.

What's new

1. Sessions list — Runtime badge (flue accented + Cloud icon, brain-box neutral) and a Spend column. Adds agent_snapshot to the web SessionSchema and a display-only flue label to lib/runtimes — deliberately kept OUT of the create picker (RUNTIMES), since flue's behavior ships in a deployed artifact (no model/key to pick).
2. Conversation viewer — distinct rendering for agent.thinking, tool results (exec.completed from brain-box AND flue's tool.result — so both render identically), turn.failed, and a body-spill affordance (content_ref/body_truncated/body_bytes → "output too large, stored in blob" instead of an empty bubble). Keeps the existing Conversation/All level filter and the live SSE tail untouched.
3. Submission health — new turns panel (GET /v3/sessions/:id/turns): per-turn state, yield_reason, timing (active_seconds), usage, and error (opaque object → one-line message). Recovery stays the existing cancel + steer and webhook redeliver — no fake "retry turn" button.
4. Per-session spend — Spend / Tokens / Events metric cards on the detail page, derived defensively from the opaque usage object.

Honest scope reconciliation (spec vs. /v3 reality)

• "Tier-2 streaming parts/deltas" does not exist in /v3. The runtime emits whole events; tiering is the level field (internal|progress|user). There's no token-delta stream to render — so "Tier-2 fidelity" = a faithful full internal-level trace, which this delivers. No view=/parts=/tier= params exist.
• No public turn re-drive/retry endpoint. Recovery is cancel + steer + webhook redeliver; the panel labels this honestly.
• No dedicated spend endpoint. Spend is derived from session.usage / turn.usage. flue meters at the gateway and reports {} here → spend renders — (correct, not a bug).

Live SSE tail

Left the working inline EventSource client as-is (resume via Last-Event-ID/?after=<seq>, seq cursor, 5s poll fallback). Did not extract a useSSE hook — extraction risked regressing the working live tail for no functional gain (per the task's guardrail).

Verification

• npm run typecheck ✅ · npm run build ✅ · preview dev server boots clean.
• npm run lint: the only 2 errors are pre-existing on main (SessionDetail.tsx synchronous setState-in-effect — the SSE reset I didn't touch; Browsers.tsx media-caption — not in this diff). Zero new lint errors.
• Preview mocks (api/mock.ts) extended (a flue session + brain-box sessions with usage, agent.thinking/exec.completed events, a turns fixture) so VITE_PREVIEW=1 npm run dev renders every new surface with no backend.
• Live prod verification is the reviewer's (localhost is wired to prod) — see run steps below.

Run against prod locally

• Fast path: OC_V3_KEY=osb_… npm run dev from web/ (Vite injects x-api-key server-side to prod /v3, per vite.config.ts), then open /sessions.
• Cookie path: plain npm run dev if your edge/session is already wired.
• Open /sessions (runtime + spend columns), a session detail (/sessions/:id — spend cards, submission-health panel, richer trace), and a running session to watch the live tail.

Base branch

Based on main — W11 is runtime-agnostic dashboard breadth that ships independently of the flue-native evergreen. If the coordinator wants it inside the flue-native unit, it's a cheap rebase (web-only, separate surface).

🤖 Generated with Claude Code

[ZIJ]

Folded into the flue-native evergreen (#489) — 1 repo = 1 flue PR. The W11 dashboard commit is now in flue-native (rides the flue PR, not a standalone main PR).