The corral (canvas view)

The corral is the default tab: every active agent in the audit herd renders as a moving node, connected by a line to each file or fault-shard it currently holds a claim on. Click any agent — on the canvas, or in the roster on the right — to filter the live console below to that agent’s lines and open its floating detail window: what it’s working on, which paths it holds, its recent activity, and the memory entries it can recall.

Anatomy of the cockpit

The corral isn’t just the canvas — it’s five regions working together, all fed by the same /api/state snapshot:

• Header stats (top center) — a running N agents · N claims count, plus a · N ⏸ suffix whenever an agent is parked awaiting approval.

• Left rail — tasks — every shard in the run’s fault-generation queue, newest work first, each row showing its phase (plant, dev-run, test-write, critic-review) and (once claimed or done) the agent’s name behind a ← arrow.

• Canvas (center) — the live force-directed graph: an agent node per active audit-role seat, a path node per file or function under audit, and a line between them for every live claim. Two mutant-generator shards claiming the same function draw two lines to one path node — that’s the shard split, not a conflict.

• Right rail — agents & findings — a roster of every active agent (labeled model/role, e.g. gemini-3.5-flash/mutant-generator) with its most recent command line, and below it the findings panel: every open survivor — a planted mutant the dev suite hasn’t killed yet — with a ⚠ N unproven counter, since a survivor only becomes a counted, proven gap once the test-writer authors a compiling test that actually kills it.

• Bottom — live console — the real command lines the herd actually ran, newest at the bottom like a tailing log, synthesized from the run’s own events, with per-agent filter chips so you can isolate one seat’s stream:

  💭 gemini-3.5-flash/mutant-generator planting goal-violating faults…
  ❯ gemini-3.5-flash/mutant-generator planted 5 goal-violating mutants ✓
  ❯ claude-sonnet-5/test-writer authored a test to close the gap
  ⚑ claude-haiku-4-5/test-critic flagged test_full_permutation — a dead range==tuple check
  verdict: CERTIFIED — killed 18/20

The agent-detail window

Click any agent to open its floating detail window: what it’s working on, which paths it holds, its recent activity, and — at the bottom — the memory entries it can recall.

The memory-recall list above isn’t scoped to that one agent — every agent’s window fetches it from the same call (s.mem.List(...) inside agentDetail in internal/ui/ui.go) that backs the Memory tab, capped at 8 entries and filtered only by the viewer’s own visibility (private vs. shared). So an agent can only ever recall what the brain’s own memory store actually holds — the shared corpus of execution-verified audit findings — never more than what the person looking at the window is themselves allowed to see.

Matrix skin

The header’s skin selector (#skinsel, defined by the SKINS table in internal/ui/web/replay-player.js) re-voices the same live audit run in a different vocabulary. Switching to matrix renames the corral to “the construct” and its tabs (proposals → anomalies, completed → archive), prefixes every agent’s displayed name with “Mr. “, swaps the ranch/grass background for falling green glyphs (renderBg’s rain case, drawn by drawRain), and reworks each role’s in-character chatter lines (ROLE_LINES.matrix) into construct-flavored quips. None of that touches the underlying run: setSkin() only rewrites DOM text and re-renders the canvas background — the same shards, claims, survivors, and live-console lines are still there underneath, just relabeled.