feat(term): full-screen terminal console for agents, sessions and cost

[Ar9av]

Adds prismor term — a curses console over the data Prismor already collects.

What it is

A two-level tree of agents → their sessions, with a live event tail scoped to whatever is selected, plus policy precedence and per-session token cost.

 >_ PRISMOR  TERM  │ Org: …  │ Policy: default  │ Agents: 35  │ prices live  │ FOLLOW
 Agents ▸ sessions  ↓last run   │  Detail
   All agents                   │  Session   d981011d-6f54-…
 ▾ ○ claude            512/229  │  Risk      70/100 · 1 findings
     6a5d59a6-e  r90 2f   — 1d  │  Immunity  active
     d981011d-6  r70 1f $426 2d │  Cost      $426.00 est · 1629 turns · claude-opus-5
 ▸ ○ codex              64/32   │  Tokens    in 3.1k · out 1.4M · cache r687.6M/w7.4M

Keys — j/k move · →/← expand/collapse · Tab switch pane · [ / ] page · Enter event detail · s sort (last run / risk / findings / cost) · f follow · v verdict filter · p policy · P pause/resume session immunity · R resume session in Claude Code · r refresh · q quit.

It is a renderer only: every number comes from prismor.runtime.store, which already fans out across workspace DBs and re-cloaks secrets on read. No new SQL. With no tty it prints a plain table, so it stays scriptable.

Cost

The event store records what an agent did but never how many tokens it took — there is no usage column and no usage payload in any event. So prismor/runtime/cost.py joins two outside sources:

• Prices — the published feed at aipricing.guru/api/pricing.json, cached to ~/.prismor/pricing-cache.json (12h TTL, 6s timeout, serves stale when offline, reports prices n/a when there is no cache at all).
• Usage — Claude Code's own transcripts, which carry message.usage per turn and are keyed by the same session ids Prismor already records.

Deliberate honesty in the numbers:

• Sessions with no transcript render —, never $0.00, and sort last under cost order. Sessions not yet priced render ·.
• Agent totals say "across N of M loaded sessions" rather than implying full history.
• The feed publishes no cache-write rate, so that component applies Anthropic's standard 1.25x input multiplier. That is the 5-minute-TTL figure; 1-hour-TTL caching bills 2x, so totals slightly understate and are labelled est.
• These are list-price API costs — on a subscription the real marginal cost differs.

Responsiveness

Measured rather than assumed:

	before	after
time to first paint	0.93s	0.11s
events per keypress	55ms	11ms
base fetch	51ms	0.7ms

• get_aggregate_stats (~900ms cold) is off the startup and draw paths entirely — the KPI panel paints instantly showing computing 24h totals… and fills in on idle (30s TTL).
• Events paginate to the visible row count instead of a fixed 200; get_events_page scales its scan with page x limit, so a screenful is 4-5x cheaper.
• Navigation is debounced 120ms — 40 keypresses with zero gap settle in 1.2s with no queued queries.
• Session pricing runs 6-per-idle-tick; the screen repaints only when something changed (it was redrawing ~16x/sec while idle).

Paging honesty

get_events_page derives total/pages from an internal window that grows with page depth (200 → 207 → 276 → 621 as you page deeper). Those are a floor, not a count, so the header shows page 3 · 23 of 207+, ] stops at a genuinely short page, and G does not jump to a fabricated last page. Session-scoped events are sliced locally from a complete list and do show exact page 1/3 of 60.

Testing

Driven through a real pty at 120x34 against live local data (35 agents, 685 sessions, 61k events): tree expand/collapse, event tail rescoping, event detail, policy overlay, verdict filter, all four sorts, paging, key-flood, idle CPU (1.9%), and the non-tty fallback.

• Pause/resume was round-tripped on a dormant session — paused → on-disk paused=True → resumed → paused=False, left in its original state.
• Claude Code resume was verified end to end: curses releases the tty via def_prog_mode() / endwin(), claude renders correctly, and reset_prog_mode() plus a repaint restores the tree. It refuses with a reason for non-Claude frameworks, a missing claude binary, or a deleted workspace.
• Secrets stay cloaked throughout — evidence panes show the @@SECRET:...@@ placeholder form.

One gap worth naming: I could not script Claude Code's exit through the test pty, so restore-after-real-claude is proven with a stub binary rather than with claude itself. Worth one manual check.

No new dependencies — stdlib curses and urllib.