Automation: Canonical dataset maintenance Automation ID: canonical-dataset-maintenance Automation memory: $CODEX_HOME/automations/canonical-dataset-maintenance/memory.md Last run: 2026-05-10T07:06:43.232Z (1778396803232) Run canonical dataset maintenance from the checked-out Alpha Research canonical datasets workspace. 1. Verify the workspace:

I’ll run the maintenance workflow from the repo, starting with the saved automation memory and workspace checks, then I’ll only start a remote run if the status output shows a dataset is available.

Workspace checks passed on codex/initial-substrate; the automation memory file is currently empty or missing. Next I’m checking remote dataset readiness and existing locks before choosing any dataset.

The remote API is unreachable from this sandbox: status retries ended with connect EPERM ...:443, so I’m stopping before starting any run, as requested. I’ll record this in the automation memory and return the blocked status table.

Remote dataset maintenance is blocked by remote_unreachable. The workspace checks passed, but npm run canonical:refresh:status could not reach https://alpharesearch.nyc after retries: connect EPERM ...:443.

I updated the automation memory at /Users/example/.codex/automations/canonical-dataset-maintenance/memory.md. There were no repository file changes, so there was nothing to commit, push, or deploy.

::inbox-item{title="Canonical maintenance blocked" summary="Remote API unreachable; retry when network allows"}

try again.

I’ll retry the same maintenance workflow now that network access is available, starting from the memory and remote status so I don’t duplicate prior blocked work.

The prior successful run already handled econ earlier today, but the last attempt was network-blocked. I’m checking the live status now to see whether econ is ready for another improvement pass or locked.

econ is ready with no active run lock, so I’m starting the required dry-run first and will inspect the rendered prompt before launching the real remote run.

The rendered prompt is the simplified mounted-volume prompt and explicitly regenerates dataset_briefing.md from current disk inventories. I’m starting the real econ improvement run now.

The run started as example-id-0005. I’m polling status now and will wait for a terminal state before trying to recover the briefing.

The run is still nonterminal and holding the econ active-run lock. The first debug snapshot shows the droplet launched but no artifacts yet, so I’m continuing to poll rather than touching docs or profiles prematurely.

The active run is still holding the lock after another status check. I’ve queued a longer poll and will use a fresh debug snapshot only once status changes or enough time passes to see new events.

The remote agent is actively working, not stuck: the latest heartbeat showed it reading existing BLS LAUS files on the mounted econ volume. I’ll keep waiting for completion and won’t start another run while this lock is active.

The dataset lock released and canonical:dataset status is back to disk_proven, but it still points at the previous inventory run id in the profile. I’m fetching the new run debug bundle to see whether the new briefing exists as an artifact or only in the transcript.

The new run failed only because it did not copy deliverables into ARTIFACT_DIR; the mounted volume did contain dataset_briefing.md, and the transcript captured it. I’m syncing that recovered briefing into the local docs and then updating the CLI-visible profile to point at this run id.

I updated the local briefing mirrors from the recovered mounted-volume briefing and posted the recovered briefing back to the remote dataset profile with the new run id. I’m doing the required readback checks now.

Readback is good now: status is disk_proven for the new run id and CLI describe returns the recovered briefing. I found a deploy:check script, so after committing I’ll run that deployment check before pushing.

The local docs and prompt records are committed. I’m running the repository’s deployment check now, then I’ll push the branch.

Retried and completed the maintenance pass for econ.

What changed:

• Updated /Users/example/Documents/Projects/alpha-datasets-canonical/docs/public-datasets/briefings/econ.md
• Updated /Users/example/Documents/Projects/alpha-datasets-canonical/docs/public-datasets/econ.mdx
• Added dry-run and real-run prompt records under /Users/example/Documents/Projects/alpha-datasets-canonical/docs/canonical-runs/econ
• Updated automation memory at /Users/example/.codex/automations/canonical-dataset-maintenance/memory.md

Verification passed:

• npm run canonical:dataset -- status --dataset-id econ
• npm run cli -- --prompt "describe dataset econ"
• npm run deploy:check

Committed and pushed: 5594272 Sync econ canonical briefing.

::inbox-item{title="Econ briefing synced" summary="Run recovered; docs/profile pushed despite artifact export failure"}

Something i dont like : Lines that just say "ZIP contents" without clarity on EXACTLY what is in there. :

Federal Reserve Economic Data extracts for UNRATE, CPIAUCSL, FEDFUNDS, DGS10, and GDP: national monthly unemployment through 2026-04, monthly CPI through 2026-03, monthly effective federal funds rates through 2026-04, 16,787 business-day 10-year Treasury yields through 2026-05-07, and quarterly nominal GDP through 2026-01; each table stores one observation per date with values in percentages or billions of dollars as published by FRED. Federal Housing Finance Agency all-transactions house price index CSVs: 10,403 quarterly state-level index observations and 83,639 quarterly metro-level observations spanning 1975Q1–2025Q4 with index values only (no confidence intervals) for U.S. states and metro areas. Treasury Fiscal Data holdings: Debt to the Penny daily balances merged across API pages (8,303 rows covering 1993-04-01–2026-05-07 in USD) and Daily Treasury Statement operating cash balance accounts (10,000 rows for 2014-06-04–2026-05-07 with closing balances in millions), plus a consolidated Treasury par yield curve CSV with daily 1-month through 30-year constant maturity yields for 2024-01-02–2026-04-30. Bureau of Labor Statistics feeds: CPI-U all items JSON via the public API (120 monthly national observations for 2015-01–2026-04) and LAUS state unemployment JSON combining 51 series with monthly seasonally adjusted rates for 2019-01–2026-03. Census Building Permits Survey county tables: March 2026 current-month permits (co2603c.txt) and cumulative year-to-date permits (co2603y.txt) delivering 3,021 county-level records each with building counts, housing units, and permit valuation by structure size. Zillow Research city-level smoothed seasonally adjusted home value index matrix (license review): 21,410 locations with monthly typical home values from 2000-01 through 2026-03; redistribution awaits legal approval. Census Bureau microdata archives preserved as delivered ZIPs: ACS 2024 1-year PUMS (person and housing files at PUMA level), CPS January 2026 basic monthly file, and AHS 2023 national PUF (housing, person, weights, and codebooks). Bureau of Economic Analysis CAINC1 ZIP archive: annual state personal income tables covering 1969–2024 stored in provider ZIP packaging alongside layout documentation. Bank for International Settlements locational banking statistics bulk archive (license review): quarterly cross-border position CSVs through 2025Q4 for reporting and counterparty countries pending legal clearance. OECD composite leading indicator SDMX JSON payload for the United States (license review) containing annual indicator values through the 2026 release governed by OECD terms. National Bureau of Economic Research business cycle chronology HTML snapshot (license review) listing U.S. recession peaks and troughs through 2020 pending legal confirmation for redistribution.

How would you improve our automation's prompt (and the prompt it kicks off the remote job with) so that it doesnt do that ? The briefings should have all the data. Show me the prompt changes.

I’ll look at the current automation prompt template and the canonical dataset prompt it renders, then I’ll tighten the wording so remote runs must inventory archive members and not hide behind package names.

The weak spot is prompts/canonical-dataset-improve-single.md: it has a much thinner briefing section than the audit/build prompts, so the remote job can summarize “ZIP archive” instead of opening/listing archive members. I’m going to add explicit archive-inspection and bullet acceptance rules there, and mirror the key archive rule into audit/build so future audits preserve it.

I added the core rule at the point where the remote job gets its instructions: archives must be opened/listed, and any briefing bullet about an archive must say what the data-bearing members contain. I’m checking the rendered prompt and existing tests now.

The rendered prompt includes the new archive rules, but an existing test checks the exact old sentence. I’m preserving that sentence and adding assertions for the new archive requirements, then I’ll remove the temporary dry-run prompt record.

The canonical prompt tests pass. I’m going to capture the final diff so I can show you the actual prompt changes, then commit, run the deploy check, and push per the repo instructions.

I changed the prompts so the remote job can’t satisfy the briefing with package-level wording.

Main prompt changes:

• Added ## Archive And Package Inspection to canonical-dataset-improve-single.md
• Required direct package inspection with zipinfo, unzip -l, tar -tf, provider manifests, codebooks, or equivalents.
• Required member-level facts in inventories: member name, sizes, detected format, row count, schema, geography/time/unit fields.
• Explicitly banned opaque briefing phrases unless the bullet also names what is inside: ZIP contents, ZIP archive, bulk archive, provider packaging, microdata files, source package.
• Added a bad/good example for the BEA CAINC1 ZIP case.
• Mirrored the archive rule into canonical-dataset-audit.md and canonical-dataset-build.md, so audit/build runs preserve the same standard.
• Added test coverage in canonical-dataset.test.ts.

The key new prompt language is:

Canonical datasets often store provider ZIP, tar, gzip, bulk, or SDMX packages. A package name is not an inventory.

For every stored archive or packaged provider payload that contains source data:

- inspect the package members directly with `zipinfo`, `unzip -l`, `tar -tf`, provider manifests, codebooks, or equivalent tools before writing the briefing;
- record member-level facts in the inventories whenever possible: member path/name, compressed/uncompressed size, detected format, row count when measurable, schema/columns when measurable, geography fields, time fields, and unit/measure fields;
- do not write briefing bullets like `ZIP contents`, `bulk archive`, `provider ZIP packaging`, `microdata archive`, or `SDMX payload` unless the same bullet also states the exact tables/files/responses inside and what each one contains.

Verification:

• npm run test:canonical
• npm run deploy:check

Committed and pushed: 7ed78b4 Require archive-level dataset briefings.

::inbox-item{title="Archive briefing prompts tightened" summary="Remote jobs must inspect package contents before summarizing"}

Improve those instructions. We should unzip it and delete the archive. Dont add prompt instructions for "too large too fully parse".

I’ll tighten the prompt from “inspect archive members” to “extract archives into explicit data files, inventory those files, then remove the archive,” and I’ll remove the escape hatch about files being too large to parse.

I found the exact lines. I’m changing all three prompt templates so archives are temporary staging artifacts: extract them into source-specific directories, inventory the extracted files, delete the archive after successful extraction, and never use “too large” as a reason to avoid parsing.

The prompt now says archives are not acceptable final storage after successful extraction, and the “too large to fully parse” language is gone. Tests pass for the canonical prompt contract.

Updated and pushed.

The prompt now requires archive extraction, not just inspection:

For every stored archive or packaged provider payload that contains source data:

- extract the package into a stable source-specific directory on the mounted dataset volume;
- preserve provider filenames and internal directory structure where useful for provenance;
- after successful extraction and inventory, delete the original archive/package file so the dataset stores explicit data files rather than opaque containers;
- inventory the extracted files directly: extracted relative path, detected format, row count, schema/columns, geography fields, time fields, and unit/measure fields;
- parse the extracted data-bearing files fully enough to support exact briefing claims; do not leave archive contents summarized only from filenames;
- if extraction fails, keep the archive only as a blocked source artifact, record the exact extraction error in quality/result files, and do not include the source in `dataset_briefing.md` as stored usable data;

I also removed the “too large to fully parse” instruction and updated audit/build prompts to match.

Verification:

• npm run test:canonical
• npm run deploy:check

Committed and pushed: 4e924cb Extract archives for canonical briefings.

::inbox-item{title="Archive extraction prompt updated" summary="Prompts now require unzip, inventory, delete archive"}