The work, in one workspace.

Fix stale and off-plan support retrieval

Diagnose why a support assistant's retrieval returns stale, off-plan or headless passages, using a recorded eval set, and fix it so it holds on held-out articles and questions.

Keep the support agent inside its budgets and tool boundaries

Repair a tool-using agent loop: enforce step and cost budgets before each call, retry only idempotent tools, send a timed-out refund to review, and run only allowed, well-formed tool calls.

Make the triage evaluation report honest numbers

Fix an evaluation harness that overstates a classifier: few-shot leakage, dropped abstentions and failures, pooled macro F1 and undefined per-label metrics.

Stop invalid extractions from reaching the shipment store

Make a model-backed extraction pipeline safe: validate output against a schema, repair only safe cases, retry provider failures within a bounded policy that honors retry-after, and route unrecoverable documents to manual review.

Add cursor pagination to the inventory list without breaking clients

A partner's inventory sync gets duplicate and missing items. Add exact cursor pagination and parameter validation without breaking clients that still page with offset and limit.

Make clinic bookings consistent under concurrent requests

Make slot booking atomic across worker processes, keep the booking and its reminder together, and keep cancellation and rebooking correct on a database that already holds double bookings.

Prevent duplicate webhook processing

A retried webhook credits a customer twice. Reproduce it, make event handling exactly once across workers and restarts, and add regression coverage.

Recover label jobs after crashes and timeouts

Label jobs get stuck after deploys and some shipments are charged for two labels. Make leases, retries and dead-lettering recover correctly, and add regression coverage.

Make the daily revenue rollup survive redeliveries, late events and reruns

Fix an event ingestion job and its backfill so each event counts once toward its UTC day, however often it is delivered and however often a range is rebuilt, and add regression tests.

Make the CLI's layered configuration follow its own rules

Fix a CLI's configuration loader so defaults, config file, environment and flags combine per setting as documented, booleans parse strictly, and every problem names its setting and source, with regression tests.

Stop a sump pump controller from short-cycling, hanging and forgetting faults

Logic harness for a firmware control loop: fix a sample ring buffer, a float switch debouncer, the watchdog kick placement and a fault latch that must survive resets, driven tick by tick.

Stop stale searches from overwriting newer results

Fix a search-as-you-type panel so results, the busy state and the error state always belong to the latest query, whatever order responses arrive in, and add deterministic regression tests.

Make bulk card moves safe across the board API and client

Fix a small board service and its client so changes stay inside the project in the path, input is validated, bulk moves are all-or-nothing in memory and on disk, and the client rolls back refused moves.

Fix tunneling and timestep drift in a 2D game simulation

Logic harness for gameplay simulation: fix a fixed-timestep accumulator and its catch-up cap, and replace a discrete overlap check with swept box collision that slides along surfaces.

Make the churn model's validation measure unseen customers

Fix a churn model's evaluation so validation holds only unseen customers chosen reproducibly, preprocessing is fit on training rows only, and AUC is correct for ties and single-class inputs, with regression tests.

Merge offline note edits without losing work

Logic harness for offline sync: merge offline note edits field by field against the last synced copy, resolve and report same-field conflicts, merge tags as a set, and keep deletes from losing edits or coming back.

Make a rolling deploy keep serving traffic and report when it stalls

Logic harness for deploy safeguards: fix rollout config validation and a rolling update controller so readiness decides availability and a stalled rollout fails at its progress deadline, all against a deterministic mock orchestrator.

Close cross-account invoice access and fix the unsafe search

Fix object-level authorization and SQL injection in a small invoicing API without breaking legitimate use, and add regression tests.

Make the rate limiter's tests deterministic and able to catch real bugs

Rewrite a flaky token bucket suite so it controls time and tests only documented behavior, make it catch seeded limiter defects, and fix the refill defect it exposes.