feat(viz): the dashboard as the project site, with the agent's session record

[1nslyn]

Summary

The dashboard is rebuilt as the project site, and the agent's discovery process is opened up: every session transcript is stored in the project and shown next to the tree.

One frontend, three sources. A static single-page app (classic scripts, no build step, hash routes) reads one record layout (record/index.json, per-run graph.json, timeline.json, sessions.json, agent_links.json, nodes/<id>.json, sessions/<sid>/turns/<k>.json) from:

• the live server, automil viz start (prints the ssh -N -L forward when it runs over SSH),
• a snapshot written by the new automil viz export --out DIR (--single-file inlines everything into one HTML page),
• a user's own local server, reached from the hosted site through the SSH tunnel: CORS is answered only for viz.cors_origins (default automil.org) and loopback origins, on the record routes and the event stream.

Views. The experiment tree in 3D (top-down layout on the page, kept lineage emphasised; a flat lineage tree one click away) beside the primary-value chart; a node drawer that states the keep/discard verdict with the graph's own helpers (paired or marginal SE, margin, companion guard), per-fold table, files changed, run-log tail, timing and the transcript turn that proposed the node; the discovery timeline (agent actions, queued/running bars per attempt, best value on one axis); the transcript reader (every turn, tool call, result and subagent, linked to the nodes it created; runs of tool-only turns fold into one line); nodes table; notes; runs page with the local-server connection; home and remote-access pages. Look: white page, black rules, a visible column grid on the hero, Manrope and JetBrains Mono, teal for data only (direction chosen from a four-option design canvas).

Session transcripts are a framework feature. automil activity ingest copies the Claude Code transcript and its subagent sidecar into automil/sessions/<id>/ on SessionEnd; the hidden automil activity store-sessions --root stores every journaled session (the campaign launcher's exit trap calls it; store_session_record.sh is removed). Open sessions are tailed from the runtime's own file and streamed as they grow.

Validation-only, one enforcement point. viz/record_graph.py projects every node and result through firewall.is_held_out_metric_key; no loader has a path under archive/<node>/certify/; run logs are served for terminal nodes only and re-redacted; secrets in tool results are masked. tests/viz/test_firewall_forged.py plants a test_auc everywhere and checks every payload and every opened path.

Also: the activity journal replay keeps which path finalized a session (hook-exporter-unreachable no longer replays as operator-close).

Deployment: site/vercel.json and scripts/site/publish.sh export the listed cell roots and deploy with the Vercel CLI; recorded runs stay out of git.

Test plan

• uv run pytest tests/ green (1815 passed, 67 skipped); coverage of automil.viz + automil.session_record at 90%
• Live server on a copied KRAS ABMIL cell in the browser: tree (3D and flat), chart, node drawer, timeline, agent transcript with folds, nodes, notes, home, runs; light and dark; phone width
• Static export served on a second port connects to the live server (CORS + SSE live); single-file export opens
• Forged-violation firewall tests; no test_/held_out/certify in an exported record
• On fir: move the five finished cells' operator/session/ records into automil/sessions/<id>/ (one-liner in the campaign README) and pull the launcher change before the next discovery job
• Create the Vercel project, attach automil.org, run scripts/site/publish.sh --deploy <cell-root>:<run-id>:<title> ...

🤖 Generated with Claude Code