A changed runbook, a fresh local Spark-X2.5 answer

A runbook says that checkout's Redis endpoint is 127.0.0.1:6379. Change the file to 127.0.0.1:6380, ask the same question, and inspect what retrieval and a local model actually return.

This Windows demo connects MemFlywheel, context-mode and Spark-X2.5 through a Node.js adapter written for the demo. context-mode ranked third in the GitHub Trending daily snapshot captured on October 12, 2026 at 19:34 Beijing time.

The input and the first retrieval

The input consists of three synthetic local sandbox runbooks: checkout Redis, reporting PostgreSQL and a search service. MemFlywheel's public writeMemoryDocument and readMemoryDocument storage functions write and verify them.

An MCP client calls context-mode's ctx_index on the directory, then ctx_search for checkout Redis endpoint. MCP means Model Context Protocol; here it carries the actual indexing and retrieval tool calls over stdio. The first result contains revision 1 and port 6379.

Landscape crop of the original CLI screenshot, sourced from rec/seed.mp4 / rec/seed.png.

Update the source, ask the same question

MemFlywheel updates checkout's file to revision 2 and port 6380. A second ctx_search uses the same query and source label, sandbox-runbooks-v1.

This call reports an auto-refreshed stale source and returns the updated runbook. The raw MCP result is retained in evidence/recorded-01/update.json. That is the concrete result of this run: a changed file produced a fresh retrieval snippet.

Landscape crop of the original CLI screenshot, sourced from rec/update.mp4 / rec/update.png.

A visible, real Spark-X2.5 request

llama.cpp, through llama-server.exe, loads ngquocvinh's community Spark-X2.5-1.7B Q4_K_M GGUF. The server uses the alias Spark-X2.5-1.7B and listens at 127.0.0.1:8099.

Node.js fetch sends an HTTP POST to /v1/chat/completions. The request contains the freshly retrieved snippet, a temperature of zero, and a maximum of 160 output tokens. The recording shows the actual tool, quantization label, requested model, HTTP 200 and returned model from that same call:

TOOL: llama.cpp / llama-server.exe
LOCAL MODEL: Spark-X2.5-1.7B
WEIGHTS: community Q4_K_M GGUF
CLIENT: Node.js fetch (real HTTP)
REQUEST model: Spark-X2.5-1.7B
HTTP 200 / actual model response
RESPONSE model: Spark-X2.5-1.7B

The actual answer is:

{"service":"checkout","revision":"2","endpoint":"127.0.0.1:6380"}

The full request and response are in evidence/recorded-01/spark-visible.json; the recording receipt preserves the real process output and exit code. Model identification is tied to the loaded file, startup parameters and actual request/response.

Landscape crop of the original CLI screenshot, sourced from rec/spark-visible.mp4 / rec/spark-visible.png.

Save the answer and read it back

MemFlywheel stores the model's answer in context/verified-answer.md. Reading through its public storage interface returns the same body; the script checks exact equality and retains the result in commit.json.

Landscape crop of the original CLI screenshot, sourced from rec/commit-visible.mp4 / rec/commit-visible.png. Capture sources, dimensions and crop coordinates are recorded in evidence-shots/manifest.json.

On this synthetic sandbox input, the local adapter completed the full sequence: update the file, refresh retrieval, infer the current endpoint, and persist the answer. The result is deliberately small and inspectable: revision 2, port 6380.

Video music: Gustav Langer, Grossmütterchen, Op.20, performed by Paul De Bra / Professor P, IMSLP #1063810, CC BY 4.0. Excerpted, loudness normalized, gain adjusted, faded and mixed under narration.