Smoke-test session 01KS7V9G893MK8RGXN2N6SBG0P — synthesis

Campaign: c-smoke-ip081-plumbing · Goal: exercise the IP-081 supervisor → sim-runner → analysis-writer → critic → close lifecycle end-to-end while sim_run_launch is still a Phase A no-op stub.

Outcome

Both sub-experiments returned stub responses in ~80 ms each — no run_ids, no metadata.json, no metrics — exactly as the brief predicted. No quantitative claims are possible from this corpus, and none are made. The plumbing is operational: the session was opened cleanly, two task IDs were issued under the session label, both reached state=finished with the params (n_agents=4 and n_agents=8) carried through verbatim, and the session-close call landed without error. Budget on every axis is essentially untouched.

Criterion-by-criterion

1. Session open returned a session_id — yes; session 01KS7V9G893MK8RGXN2N6SBG0P was created on the prior session-open call and was already in running state at the start of this supervisor turn.
2. ≥2 sub-experiments produced structured outcomes — yes; task IDs 01KS7VS96WTN4QH2V5JS42NCWY (n_agents=4) and 01KS7VSDRH71652W3G10J1VFR9 (n_agents=8) both terminated cleanly with the rationale and session_id labels propagated.
3. Session close completed — this artefact itself is the evidence.

Caveats and follow-ups

The sim-runner / analysis-writer / campaign-critic subagent fan-out described in _agents/campaign-supervisor.md Steps 4–6.5 was inlined by the supervisor — those vault prompts are loaded by the in-repo ReactAgent SDK adapter, not by Claude Code's native Agent tool, so spawning them as Claude Code subagent_types would have failed. The lifecycle was exercised in full, but the subagent-isolation boundary was not. When sim_run_launch is wired through Phase B (real Docker runs against jupedsim-runner / pedpy-analyser) and the campaign is run via monad-knowledge sim campaign run … --supervisor react-agent, both gaps close in one step.

No thesis-chain footnote applies; this campaign is platform regression, not science.

Claim audit

The synthesis makes exactly one substantive claim: "the IP-081 plumbing — session open → run launch (with session label) → run terminate → session close — is operational under the Phase A stub." That claim is supported by direct artefact evidence: two sim_run_status payloads reaching state=finished with session_id propagated, plus the close-call return itself. The synthesis does not claim parameter sensitivity, gate behaviour, lift, or any metric — appropriately, given the stub.

Weak claims

• The synthesis notes runs_attached: 0 on the session status pre-close but does not flag whether this is a real plumbing finding (the close call's attached-runs counter should remain 0 in Phase A and grow to 2 in Phase B+). Worth a single sentence in the next smoke session once sim_run_launch is wired to the real SimulationRunner, so this becomes a regression sentinel.
• "Plumbing operational" is asserted globally; the supervisor did not actually exercise sim_run_cancel, sim_artefact_signed_url, or sim_fetch_s3. Those are out of scope for this smoke test but should be named explicitly in a future v2 of the brief if the platform wants broader regression coverage.

Honest scope

The subagent-isolation boundary was not exercised — sim-runner / analysis-writer / campaign-critic were inlined because they are not Claude Code subagent_types (only ReactAgent SDK-side vault prompts). The synthesis acknowledges this explicitly, which is the right disclosure. The lifecycle as a whole is operational via the MCP tools; whether it is also operational via the SDK subagent fan-out is a separate assertion requiring a CI run of sim-campaign.yml to validate.

Verdict

Agree with synthesis. Severity: low.