PIMP MY IDEBack to dispatches
October 11, 2026 / distributed systems / test craft

A seed is not the replay.

A failing simulation becomes useful when the code, settings, seed, action limit, injected faults, and checked invariant travel together.

Save the whole execution contract. A seed can reproduce one random choice stream only when every input around it stays fixed.

Determinism has inputs

Celld runs production event-handling code inside a simulator that chooses the next request, asynchronous task, storage result, clock advance, and fault. The project says the same code, settings, and seed produce the same sequence and intermediate states.[1]

That sentence names the contract. The seed does not identify the code or the settings. It also does not describe the invariant that failed. A bug report that saves only the seed leaves the next engineer to guess the rest.

The replay unit is the execution contract, not the lucky number.

Separate choosing from doing

Celld separates the code that selects the next event from the code that handles it. The simulator can delay a storage write, make it fail, advance simulated time, or run another task before a pending request completes. A checker tests invariants after each action.[1]

One generated order exposed an alarm race. A delayed cleanup removed a wake entry after a newer alarm had reused it. Both client operations had returned success, but no wake entry remained to reactivate the cell. The invariant caught the mismatch between the alarm stored in SQLite and the missing wake entry.[1]

The reusable move is architectural. Put scheduling, fault choice, and time behind controllable interfaces. Keep the production handler on the other side. Then save every choice needed to replay the failure.

One process can contain a bad week

FoundationDB's Flow runtime supports production execution and deterministic simulated execution. Its simulator runs an entire cluster in one single-threaded process. It models machines, drives, networks, communication delays, restarts, and datacenter failures.[2]

A single-threaded simulator does not claim production has one thread. It gives the test harness control over execution order. FoundationDB says that determinism allows exact repeatability and controlled experiments. Different tests and seeds still matter because one repeatable schedule covers one route through a much larger state space.[2]

Agent throughput raises the receipt bar

Cockroach Labs describes a coding-agent workflow with separate planning, implementation, review, and discharge roles. GitHub issue labels carry workflow state. Structured notes carry each stage's record. The process requires a reviewed plan before code and sends rejected work back through an explicit loop.[3]

That workflow is evidence about Cockroach Labs' own migration tooling. It does not prove that the same roles will work in another repository. It does point at a practical pressure. When patches arrive faster, failures need better records. A reviewer should be able to hand a simulation failure back with the exact revision, settings, seed, action budget, fault set, invariant, and observed step.

Ship the failure as a cassette

Use a replay packet with fields that can be checked rather than a prose note that says "flaky under seed 184467." At minimum, record:

  1. The source revision and simulator revision.
  2. The complete settings or a digest of an immutable settings file.
  3. The seed, random-number generator version, and action limit.
  4. The permitted requests, faults, delays, and clock behavior.
  5. The invariant name and the first step where it failed.
  6. The replay command and the result after the proposed fix.

Keep the first failing packet immutable. A fixed run is a second record. If the fix changes the settings, action budget, or invariant, say so instead of calling the new run a replay.

Interactive handoff fixture

Determinism replay deck

Traditional purpose: a bug form, run configuration, checklist, and copyable handoff. This deck keeps them on one transport and labels the result as a template until a real run fills the evidence fields.

CodeSettingsSeedInvariantReplay

Load the cassette

Enter the known inputs. Select the records that the real failure packet will attach.

Maximum actions500
10 actions1000 actions
Choose replay packet attachments
REPLAY PACKET OPEN0 of 4 records selected

Execution contract

The transport shows packet completeness. It does not run a simulator.

PACKET INPUTS OPENNo simulation has run. Add the missing identifiers and select the records that a real failure handoff will attach.
Sources read / October 11, 2026 UTC

Source log

Open the evidence boundary
  1. Celld, "Deterministic simulation testing in celld", published October 2 and read October 11, 2026. This supports the production-handler boundary, controlled scheduling, simulated time, storage faults, seed contract, action limits, invariant checks, and alarm-race example. Celld says its simulator is still under development and is not in the public repository.
  2. FoundationDB 8.0 documentation, "Simulation and Testing", read October 11, 2026. This supports Flow's production and simulated execution modes, single-threaded deterministic cluster simulation, modeled failure types, and repeatability claim.
  3. Cockroach Labs, "Five months treating bugs like patients and coding agents like a medical team", published October 9 and read October 11, 2026. This supports the role split, GitHub-label state machine, structured notes, reviewed-plan gate, and explicit rework loop. Its reported outcomes apply to that team's MOLT workflow.
  4. Hacker News discussion for the Celld article, item 50024805, verified through the HN search API and read October 11, 2026. The thread had one comment when checked. We used it only as a current-interest signal, not as technical evidence.

Evidence boundary: Pimp My IDE read the named pages. We did not access Celld's private simulator, run FoundationDB simulation, or reproduce the alarm race. The replay deck writes a handoff template. It does not execute code, validate a seed, or attach real evidence.