[EiffL]

Why

Opening an output in lc view showed Status unavailable, and the figure cards carried no freshness marker, even on a fully current project. The page filled astra-ui's renderProvenance slot with the recorded run but never its getOutputStatus slot (or the provenance panel's status), so astra-ui fell back to its unknown label.

What

• One route, GET /lightcone/api/status, answered by materialize.status — the same call lc status makes — as {outputs: {"<universe>/<output id>": {state, detail}}}, run off the event loop since it hashes declared inputs. A project the engine cannot read is a 422 carrying the engine's message. Token-gated like every other route.
• The page asks it after every resolution (whatever changed on disk may have changed a verdict) and hands the lookup to Inventory and the open output's provenance panel. A failed request shows no state rather than an error.
• The browser still holds no copy of the classification rule, which is how jupyterlab-lightcone fills the same slot. CLAUDE.md's host-slot list is updated accordingly.

Tests

• test_an_outputs_state_is_the_engines: a real project served, the route's JSON equals materialize.status(root) in-process.
• test_a_project_the_engine_cannot_read_has_no_state: 422 on a directory that is not an lc project.
• /lightcone/api/status added to test_every_route_needs_the_token.

Verified by hand on a materialized project: the dialog reads Status: Current with the run's time and revision.

🤖 Generated with Claude Code

[github-actions]

✅ Eval

Metric	Value
Outputs check	success
Agent run	success
Turns	8
Tool calls	6
Cost	$0.11
Agent wall time	0m42s
Model	claude-sonnet-5-5

lc status

  mode:    direct
  sandbox: landlock (fs: declared, network: allowed)
  crate:   up to date with the outputs

  · current  baseline/best_fit        ebed420
  · current  baseline/hubble_diagram  ebed420
  · current  baseline/residuals       ebed420

3 current

Confusion & pain points (Claude analysis)

Confusion & pain points

• The run was essentially clean. There were no errored tool results, and the agent built all three outputs, validated the spec, and passed --check in 6 tool calls.

• The agent never saw the output of its heaviest step and did not verify it. It ran spec editing, uv add, validation, git commit and lc compute launch in one chained command piped through tail -3. It never confirmed the commit succeeded or that the astra.yaml text replacement matched, and the final git status --short came back empty from view. Cause: tail hides failures, and the chain continues after a failed step. The final message admits that it did not confirm the working tree was clean or that ro-crate-metadata.json was committed. A materialize that exits 0 doesn't show the tree is clean, so lc materialize could print a one-line "tree clean / crate committed" confirmation.

• The agent edited astra.yaml with a blind Python str.replace. It matched literal recipe:\n command: python scripts/... text to add format, inputs, decisions and the command. If the scaffold's whitespace had differed, nothing would have changed and no error would have been raised. Only the later astra validate pass shows it worked. This hints that the scaffold doesn't make completing a partial spec easy, for example with an astra command that fills in the missing output fields.

• The agent had to discover format: and the license requirement itself. Every output needs format: because lc has nowhere to write the file without it. A [project].license is also needed to get the RO-Crate. The agent added both correctly, but the license edit used a fragile sed that deleted any existing license line and then inserted after version. Suggestion: the scaffold's pyproject.toml could include a commented license placeholder, and lc materialize could say when it skipped the crate for lack of one.

• Cluster lifecycle is an extra manual step. The agent had to run lc compute launch --wait, then lc materialize local, then lc compute down local. Forgetting the last step leaves an allocation running. It handled all three correctly this time, but the sequence is easy to get wrong for an agent that has never seen it.

Full trace: agent-trace artifact on this run.

[EiffL]

lgtm