Pome run task (Skill 4)

You are the coach: you talk to the builder and to the Pome control MCP (mcp.pome.sh). The examinee is the sandbox clone Skill 0 (pome-intake) registered, launched here for real against live twins. This skill runs one already-verified task and scores it from the twin tape. Verify the seed first (pome-verify-seed) — this skill does not re-check fairness.

Only one step here is examinee-runtime-specific: launching the examinee. Minting, finalizing, scoring, reporting, and the fix loop are runtime-agnostic (ADR-018). The launch policy — always_allow, closed-book web tools, memory snapshot-clone, the network clamp — is owned by the examinee_launch spec, not by this prose: you execute the spec, you do not restate it.

If the mcp__pome__* tools are missing, the MCP isn't connected: ask the user to connect and authenticate it (interactive OAuth — needs a human in a browser) instead of probing the endpoint.

SENSITIVE — the bearer. run_task returns agent_token, a session-scoped JWT that is the bearer for every twin URL — a live credential. It exists for one purpose: handing the examinee its twin authentication at launch (§2). Pass it straight into the launcher's env or vault and then let go of it — never write it to disk, into a task, or a log, never repeat it back to the builder, and never keep it around "for later". Nothing downstream needs it: finalize_run derives the bearer from session_id server-side. It dies at expires_at regardless.

1. Mint the run

Mint a grp_-prefixed group_id now and reuse it for every trial of this attempt — the baseline and any pre-fix flaky retries — so they aggregate as one exam (aggregation keys on (group_id, task); never reuse a group_id across different tasks). A post-fix rerun is the exception: it opens a new group_id and links back to the baseline via baseline_group_id (see §5). Then call run_task(task_id, agent_id, agent_version, group_id) (the agent_id from intake). It seeds live twin sandboxes and returns session_id, expires_at, agent_token, examinee_task (the prompt + twins the examinee sees — no criteria), and examinee_launch (the full launch spec).

Before launch, show eval_cost to the builder and ask for confirmation. Do not launch the examinee until the builder confirms the cost.

agent_version on every run. Read it from the manifest's agent.version field. If the builder declares none, ask for one before the first run rather than sending nothing. It is the label the run declares itself to be, and it is what keeps one version's trials from being averaged with another's — the dashboard partitions run-sets by (agent, task, agent_version). Never auto-bump it: it changes when the builder changes the examinee (§5), and a version that moves on its own would split one exam into two run-sets of one.

• Trials-of-N (the batch form) — when the task's ## Config sets runs: N (a flakiness budget), don't hand-loop run_task: call run_trials(n, task_id, agent_version, group_id), the batch form that provisions all N trials up front under one shared group_id and returns a trials[] array, each with its own session_id + agent_token. You still launch and finalize_run each trial (all N sandboxes share the quota — launch + finalize promptly to free slots). Pass-rate is judged here, not by the platform: once the trials finalize, list_runs(group_id) gives the cross-trial view — compute the fraction passed and compare it to the task's passThreshold (default 100%).
• twins not enabled (HTTP 400) — the agent's allowlist is missing a twin the task needs. Heal with one additive register_agent(name, twins:[…]) (it merges, never removes), then re-run. Do not re-intake the scope.
• Everything the examinee needs is inside examinee_launch. Read the Runtime line from the intake report (or examinee_launch.transport) to pick the launcher below.

2. Launch the examinee (the one runtime-specific step)

Dispatch on the runtime and hand off to the matching launcher, which assembles the examinee faithfully from examinee_launch, starts it on the kickoff task, and watches for idle:

• Claude managed agent (transport: "mcp") → Anthropic's Managed Agents cloud via the ant CLI. Recipe: references/launch-managed-agent.md.
• Anything else (transport: "rest") → the REST path (rest_urls + env). Recipe: references/launch-rest.md — which preflights the wiring (config → twin reachable → routing → egress floor, the pome doctor checks) before it launches, so a mis-wired examinee never runs against a live API instead of the twin. Dispatch honestly — a non-Claude examinee on Managed Agents runs as Claude, testing nothing.

3. Finalize the instant it idles

The twin tape lives in the running session's sandbox. finalize_run captures final state + events off the still-live twins; once the session leaves ready/running the sandbox is torn down and the tape is gone — it errors, and the run is unrecoverable. So the moment the launcher reports the examinee idle (done / awaiting-input with no more tool calls coming), call finalize_run(session_id) immediately — before any cleanup, before narrating anything. It scores synchronously against the pulled tape and returns { run_id, score, judge_model, dashboard_url }. One evaluation per run.

4. Narrate the report

get_report(run_id) returns the run markdown. Narrate, don't dump: the Score /100, the criteria table (each row's Kind = code/model, Status = passed/failed/unmatched, Reason), Provenance (a live twin-pull run is hosted — say so, it means Pome watched the work, not the agent self-reporting), and the dashboard link on app.pome.sh. An unmatched criterion binds to no declared check, so it was never graded — that is an authoring defect, not a failing grade. Route it back to pome-author-task, which re-authors it from list_checks rather than rewording it.

5. Fix loop (re-run only what failed, show the delta)

A green run is done. On a failure, the report's ## Handoff (fix prompt) section is the driver: it names what the agent did wrong. Hand it to the builder, they edit the examinee's prompt, then re-run.

The one thing you never do: make the exam easier to pass. Every fix goes into the examinee's prompt — never into the task. Do not weaken or delete a criterion, lower a passThreshold, loosen a [code] predicate, or edit/remove the seed or its expected end-state to turn a red run green. That is the "vibe-coder" failure DeepEval names — gaming the metric by rewriting the test instead of fixing the work — and it silently destroys the exam: a task that no longer discriminates a working agent from a broken one grades nothing, so a green it produces is worthless. If a criterion is genuinely wrong (unfair, unmatched, or mis-specified), that is not a fix-loop edit — stop the loop and route it back to pome-author-task / pome-verify-seed, where any criterion or seed change is re-verified as a fair exam before it counts. Then:

1. Re-run only the failed tasks as a fresh run-set — one run_task (or run_trials for a flaky task) each against the same agent_id. A post-fix rerun mints a new group_id (omit it and run_trials mints one) and passes the failing run's group_id as baseline_group_id — the report's ## Rerun after fixing section pre-fills both. The rerun is its own run-set linked back to the baseline; the dashboard pairs them and shows the fail→green delta. Do not reuse the baseline's group_id — that merges baseline+green into one aggregate and destroys the split. (A pre-fix flaky retry — same examinee, no edit — still shares the group_id; only a post-fix rerun opens a new one and links back with baseline_group_id.)

   Bump agent_version too, and say so to the builder. The edited prompt is a different agent, so the rerun declares a different version — have the builder set agent.version in the Pome manifest, or agree one with them and pass it. A fresh group_id alone is not enough: the reliability page also partitions the implicit run-set by declared version, and the verdict strip asserts "same agent, same prompt" over a run-set. Rerun the fix as v1 and the platform is being told the failure and the fix are the same agent — the spread it then reports is the fix working, mislabelled as unreliability.

2. There is no delta field in the report — compute it: pull get_report for the baseline run and the rerun, diff the Status column per criterion, and report the flips ("leaked to #general: failed → passed"). Baseline and rerun are now separate groups paired by baseline_group_id; list_runs(group_id) gives each run-set's view.

3. Repeat until every re-run is green (or the builder accepts the behavior). Only failed tasks re-run; the green ones are not re-billed.

Report

End with: the task name and run_id, Score /100, the criteria table (criterion · kind · status · reason), provenance, and the app.pome.sh link. On a fix-loop run, add the per-criterion delta and name the prompt edit that moved it.