CedarCopilot wants to merge 4 commits into staging from audit/warp-data-audit
Live on prod, no production signal yet
Strict CloudWatch fallback found 5 prod failure log lines for [processIncomingSlackChannelData] Error processing channel (No valid connection tokens found for user <id>) in the last 24h while checking process, deltas, readme, nextsteps, usecases.
Strict CloudWatch fallback found 5 prod failure log lines for [processIncomingSlackChannelData] Error processing channel (No valid connection tokens found for user <id>) in the last 24h while checking /server, /seed-org-background-fields.ts, applied, customfielddefinitions.
Strict CloudWatch fallback found 5 prod failure log lines for [processIncomingSlackChannelData] Error processing channel (No valid connection tokens found for user <id>) in the last 24h while checking fieldname, process, deltas, readme, nextsteps.
Strict CloudWatch fallback found 5 prod failure log lines for [processIncomingSlackChannelData] Error processing channel (No valid connection tokens found for user <id>) in the last 24h while checking process, deltas, readme, nextsteps, usecases.
Strict CloudWatch fallback found 5 prod failure log lines for [processIncomingSlackChannelData] Error processing channel (No valid connection tokens found for user <id>) in the last 24h while checking skipinstance, judgment, process, deltas, readme.
Libra has verdicts on 0 of 9 tracked behaviors on prod; 9 are still being checked. Libra has 5 low-confidence degraded verdicts it is still confirming. Libra checks hourly for 3 days after each deploy.
The Warp new-business data audit (Kendall / Will / John, Feb,Aug 2026): the requirements + data-problems record, the regeneration tools, backing data, and the process-map artifact.
Depends on nothing; sibling to #2592 (the two stage-inference fixes this audit surfaced), which is separate so those logic changes get read on their own.
Ground truth is Salesforce IsWon/IsClosed, with the February Attio migration import excluded , those deals were created and closed in the same batch and inflate every rate.
| Script | What it does |
|---|---|
scripts/tpuf-parity.ts | Read-only Turbopuffer↔Postgres parity for an org. Instance plane compared by id, not by count, so an offsetting pair of errors can't hide; judgment plane by count (snapshots are keyed by snapshot id , nothing to join on). |
scripts/tpuf-reconcile.ts | Targeted repair , deletes orphaned embeddings, embeds missing rows, scoped to a conversation set, dry-run by default. |
These exist because of a structural bug worth knowing about: every taxonomy rewrite deletes and re-inserts, minting new Postgres ids, so any generation whose paired Turbopuffer delete failed is orphaned rather than overwritten. It survives carrying superseded content, and semantic reads keep returning it.
On Warp that was 2,694 orphans across 188 conversations, 87% still holding pre-canonicalization values against 15.4% in Postgres. Nothing in the write path detects it. Now reconciled , instance 12,392 = 12,392, judgment 1,862 = 1,862, 0 missing / 0 stale.
reconcile-turbopuffer.ts was not used: it deletes every instance embedding in the namespace and re-embeds the whole org (~12k needless embedding calls, org-wide blast radius) and is hardcoded to the Pirros org.
canonicalize-judgment.mjs , the instance-plane canonicalizer globs rows_*.json only, so a customer vocab living on a versioned field was invisible to it and silently never canonicalized. Real case: qual.business_outcome.value_driver, 235 assessments across 153 deals, 100% other, reported as "0 moments" because the harvest was looking in the wrong plane.extract-workflow.js , skipInstance flag, so the judgment plane can be re-run alone instead of redoing a completed instance extraction.docs/audits/warp/tools/ regenerates the artifact end to end. audit-quotes.py sweeps every shipped quote and play exhibit (4,019 sides) for speaker mis-attribution, cross-plane duplicates, filler, fragments and truncation, and exits non-zero on a blocking class , currently zero in all seven classes. scratch/ is excluded per the data-audit skill's placement rule (recomputable intermediates and raw dumps).
The artifact's brand palette is inferred, not sourced. --accent:#4338CA and the logo mark were designed to fit Warp's positioning; warp.co serves its CSS externally so I couldn't read the real values. Everything brand-related is isolated in tools/warp-shell.html , swap the tokens and rebuild.
🤖 Generated with [Claude Code](https:/
Libra has not measured any production surfaces for this change yet.