[MaxGhenis]

Builds on #110 (base uk-data-issue-routing). Retarget to main once #110 merges.

policyengine-api's economy routes reject any query parameter they don't declare: StrictQueryParameters sets extra="forbid" in policyengine_api/query_parameters.py. microsim-runner's US call sent dataset=enhanced_cps, as did its budget-window helper and model-corroborator's mirror call, so each got HTTP 400 "dataset: Extra inputs are not permitted". The routes build every request with dataset="default" and run it on the one dataset /{country}/metadata lists under economy_options.datasets. On 2026-11-08 the metadata listed Microcosm's populace_us_2024 at populace-us-2024-spm-20260915 for the US and microcosm_uk_2024_25 for the UK.

Plan (WIP)

• Give up on a budget-window batch that keeps computing, and fall back to single-year calls
• Independent review agrees at head
• Rebase onto main and retarget after Route UK data issues to the repo that builds the run's dataset #110 merges
• Merge on green checks

Changes

• microsim-runner
  • Sends no dataset on the annual or budget-window route, and lists the parameters each route takes.
  • run_dataset replaces the advertised-dataset table and the "populace" in data_version check. It names the dataset from the hf:// URI in the result's policyengine_bundle the way policyengine.py's dataset_logical_name does. It fills a missing data version from the URI's @ revision, and anything else the bundle leaves out from the metadata default.
  • Takes the local path for a named dataset whatever mode says, and passes the name to pe.us.ensure_datasets.
  • On the budget-window route, the runner reads the result policyengine-api relays from the gateway's BudgetWindowResult: annualImpacts rows and a totals row, in dollars, with no bundle.
  • run_horizon takes a region instead of reusing the country, raises on an error status instead of returning None, and sets a timeout on each request. It gives up once a batch keeps computing past 45 minutes, and the runner then falls back to single-year calls. On 2026-11-08 a live two-year US window was still scoring 2026 two hours after the first request, while the same reform's single-year run for 2026 took about three minutes.
  • In step 3 the runner scores multi-year runs on the budget-window endpoint. The response example drops reform_net_income, which the live result lacks.
• model-corroborator: the mirror call sends no dataset.
• reform-comparator: records dataset, data_version and model_version from the runner's output, and notes a run whose dataset_matches_default comes back false.
• policyengine-api and policyengine-app skills: list the economy query parameters, and describe the budget-window route and its status envelope.
• test_dataset_naming.py: replaces the advertised-name test with property, differential and doc-scan tests. With PE_SKILLS_LIVE_API=1, it also checks the metadata snapshots and the 400 against the live API.
• test_budget_window_polling.py (new): runs run_horizon against scripted answers and a fake clock.
• test_no_stale_references.py: drops the runner's frozen "populace" in data_version span, bans that sample on the runner, and scans the dataset test, which names datasets by their frozen ids.

Invariants

Both test files exec the functions straight from microsim-runner.md, so the agent's code and its tests can't drift apart.

run_dataset, for any bundle whose dataset holds an hf:// URI, a bare name, default or nothing:

• it returns a dataset name: never default, a path, an @revision or an .h5 suffix;
• an empty bundle, or one that reads default with no data version, gives the metadata default, and dataset_matches_default holds;
• the bundle's own dataset and data version win over the metadata;
• a missing data version comes from the URI's @revision before the metadata;
• dataset_matches_default holds if and only if both values equal the default entry's.

For any non-empty dataset string other than default, run_dataset returns the name policyengine.py's dataset_logical_name returns. A differential test calls the installed package's function, or a copy of its 6.16.0 body when policyengine isn't installed.

run_horizon, for any run of computing answers, any ending and any deadline:

• it returns the result of the first ok answer, never None;
• any other status raises RuntimeError with the API's reason;
• a batch that keeps computing raises TimeoutError at the first poll at or after the deadline, never earlier, and a terminal answer on that poll still wins;
• it sends no request after a terminal answer;
• every request carries the region, start year and window it got, plus a timeout, and no dataset.

test_no_agent_or_skill_sends_dataset_to_the_api reads every Markdown file under skills/, targets/ and docs/. It flags dataset in a query string, or as a key or keyword inside a code block that calls /economy/, and two more tests pin what it must catch and allow.

Verification

• uv run pytest --ignore=tests/test_skill_examples.py: 297 passed, 6 skipped (the opt-in tests), 1 xfailed.
• Each of 14 mutants of run_horizon and run_dataset fails a test, including ones that drop the deadline, check it before the status, send dataset, ignore the URI's revision or leave it on the name.
• With PE_SKILLS_LIVE_API=1, both metadata snapshots match api.policyengine.org, and the US annual, US budget-window and UK annual routes each answer dataset with HTTP 400 "dataset: Extra inputs are not permitted".
• With policyengine 6.16.0 installed, run_dataset and the pinned formula agree with dataset_logical_name on generated strings.
• A live US annual run (policy 97853 over 2, region us, 2026) returned a bundle naming hf://policyengine/populace-us/populace_us_2024.h5@populace-us-2024-spm-20260915, which the check maps to populace_us_2024 and the metadata default.
• scripts/build_claude_wrapper.py builds, and the built runner carries the check and the deadline.

🤖 Generated with Claude Code