Browser Casefile

The browser is an inspector and operator surface for persisted cases. Generated/custom workflow execution works only for packs explicitly registered when the API process starts.

For existing agents wrapped with cover.case(...), run the agent script first. The persisted governed case appears in the same case list and uses the same execution report. Repeated runs of the same cover.case("name", ...) appear as a linked run series so you can move from one case run to the previous or next run without treating each UUID as unrelated.

Generated Workflow Path

Create a generated pack and prepare the database:

casegraph init intake_review --output .casegraph/intake_review_pack --json
casegraph db upgrade --json

From this repo checkout, prefix product commands with uv run --all-groups --python 3.13.

Start the API with the generated pack registered:

CASEGRAPH_PACK_ROOTS=.casegraph/intake_review_pack \
  uv run --all-groups --python 3.13 uvicorn casegraph.api.app:app --reload

Start the web app:

pnpm install --frozen-lockfile
pnpm --filter @casegraph/web dev

Open the browser:

http://127.0.0.1:5173

Click the registered intake_review@v1 workflow, then click Open demo case. The browser can then run workers, generate a proposal, preview, assess policy, record approval, plan commit, execute commit, and inspect replay. The same path stays local/fake and provider-free.

If a generated/custom case is not registered, the browser keeps it in read-only inspection mode and shows a CASEGRAPH_PACK_ROOTS fix.

Existing Agent Path

A script that uses this shape persists a browser-inspectable case:

from casegraph import cover

with cover.case("incident_triage", goal="Resolve the incident") as case:
    result = agent.run_sync(prompt)

After the script prints case.outcome.inspect_command, start the API and web app and select the latest case.

What To Inspect

Panel	What it shows
Case header	Intent, engine, status, counts, and run-series metadata when present.
Execution report	Worker evidence, proposed actions, preview, policy, commit, replay, and recent events.
Frontier	Remaining or resolved gaps.
Evidence	Source refs, tool calls, observations, claims, and worker runs.
Prompts and model runs	Rendered prompts and opt-in model-call provenance.
Proposal and preview	Suggested actions and expected before/after change.
Policy and approvals	Readiness, blocked/auto status, and operator decisions.
Commits	Plan, attempt, receipt, verification, and compensation records.
Replay and timeline	Event history and in_sync replay status.

Support Reference Case

For the larger support escalation reference/demo pack:

uv run --all-groups --python 3.13 casegraph demo seed --json

That case keeps the browser operator buttons for the bundled support reference flow. It does not mutate real external systems.