Diagnostic
$900
- One target surface reviewed.
- One reduced repro or risk report when a bug is found.
- Short mitigation plan if no clean fix is feasible.
Proof-led compatibility pass for database, sync, and query teams
Turn one risky query surface into merged fixes and regression tests.
I run a focused hardening sprint against one narrow surface: SQL compatibility, query planning, schema rewrites, sync edge cases, SDK adapters, or import/export paths. The deliverable is a reduced repro, a failing test, a fix PR when feasible, and a maintainer-grade handoff.
Recent paid-sprint trace
4 merged PRs
Buyer-funded scopePowerSync paid for a focused SQLite compatibility sprint.
Paid
Four accepted fixesPRs #644, #645, #646, and #647 were merged after maintainer review.
Reduce
Review-safe finishFollow-up fixes landed only when maintainers requested them, with the paid scope kept clean.
Fix
Current proof queueInstantDB merged, with RxDB, Dexie, Prebid salesagent, Electric, and Convex PRs showing the same reduced-repro style.
Ship
Payment boundaryNo checkout link is sent until scope and start intent are confirmed.
Clean
Fixed-scope sprint
Best fit when you have a risky compatibility surface but not enough time for broad QA.
48-hour sprint
$2,000
- Focused fuzzing or differential test harness pass.
- Reduced repro and failing test.
- Patch PR when the fix is inside a small, safe scope.
5-day hardening
$5,000
- Everything in the sprint.
- Two to three adjacent bug classes or adapter surfaces.
- Repro suite, maintainer notes, and follow-up hardening checklist.
Shipped proof
These are real public artifacts, not screenshots of intent.
PowerSync — 8 PRs (4 merged, 4 open Jun 6-7)
Buyer-funded Sprint #1 merged four LIKE/range/cast/json_each fixes (#644-#647); follow-on weekend opened JOIN-loud-error #662, ASCII upper/lower #663, length() code points #664, substring() code points #665.
TanStack DB PR #1574 (Jun 7)
ASCII case fold in upper/lower/ilike so the JS evaluator agrees with the source DB on the German sharp s, ligatures, and Turkish dotted-i. All 99 evaluator tests pass.
TanStack/cli PR #467 (Jun 7)
Documents the four sync-stream SQL footguns in the PowerSync add-on starter so generated apps know which queries silently drop rows.
silentdrop (MIT, npm)
Open-source correctness checker that runs the four divergence classes (LIKE-newline, LIKE-escape, case-fold-length-change, non-BMP ordering) against any JS database query layer. Same shape as the PowerSync / TanStack DB / Zero / InstantDB / Electric / Dexie / RxDB / Autumn fixes.
silentdrop-llm (MIT, npm)
Same lens applied to LLM/agent JSON outputs — missing required fields, enum drift, hallucinated IDs, claim-vs-list mismatches. Runtime guard before downstream code commits to wrong state.
InstantDB PR #2714
Merged client-side $like/$ilike newline compatibility fix; maintainer said they would publish a new SDK version.
RxDB PR #8601
Query-planner regression for out-of-schema numeric bounds, with CI blocker fixed after review.
WatermelonDB PR #1968
LokiJS adapter Q.like regex-escaping fix for SQLite-compatible matching.
Dexie PR #2306
equalsIgnoreCase() fallback for length-changing Unicode case folds that would otherwise silently drop rows.
Prebid salesagent PR #1380
AdCP v3 pricing-field fix for authenticated buyers seeing the anonymous-pricing message.
Electric PR #4437
Postgres LIKE compatibility fixes with function-level repros and green code checks.
Turso issue #7146
Reduced invalid CTE UPDATE repro: SQLite errors and leaves data unchanged, while Turso accepts the statement and mutates a row.
Turso PR #7144
Fixes ALTER TABLE RENAME COLUMN when a TEMP trigger targets a main table. Includes the regression and exact fuzzer seed validation.
Turso PR #7143
Fixes a subquery scheduling panic around aggregate arguments and FILTER predicates.
Context OS
Installable GitHub Action for agent-context cost and failure-surface checks.
Agent cost leak checker
Public repo scanner showing how I package technical findings into operator-readable outputs.
Send one risky surface.
Good inputs: a repo link, a failing compatibility area, a query feature, an SDK adapter, or one engine behavior you do not fully trust.