Micro-context S1 — 100 contexts through one real endpoint

Verdict

PASS at four physical workers: 100/100 logical contexts retired through the native OpenAI-compatible fak endpoint on the sanctioned GCP fak-realmodel node. This advances the program from synthetic S0 scheduling to a real model/provider seam without starting 100 full agent harnesses.

The endpoint used the node’s qwen2.5-0.5b-cpu fak-native model path. The node has an L4, but this run deliberately labels itself CPU path on an L4 node; it is not a CUDA throughput claim. The separately exposed CUDA endpoint stopped returning completion responses during preflight and therefore supplied no successful witness.

Reproduction

go run ./cmd/microcontextdemo \
  -selfcheck -contexts 100 -workers 4 \
  -endpoint http://<sanctioned-node>:8081 \
  -model qwen2.5-0.5b-cpu \
  -provider fak-native-openai-compatible \
  -hardware "GCP fak-realmodel CPU path on L4 node" \
  -request-timeout 5m

go test ./cmd/microcontextdemo -run 'TestSpine|TestOpenAIEndpoint'

The checked-in public artifact omits the endpoint address. The node identity and command are sufficient for an authorized operator to reproduce through the sanctioned runbook.

Observed pass

Source: s1-gcp-realendpoint-workers4-pass-2026-08-06.json.

Metric	Observed
Logical contexts	100
Physical workers	4
Completed / failed	100 / 0
Shared base installs	1
Wall time	94.725 s
TTFT p50 / p95	3.852 s / 3.939 s
Prompt / completion tokens	6,190 / 800
Aggregate prompt tokens / wall second	65.35
Aggregate decode tokens / wall second	8.45
Useful context completions / second	1.056

prompt_tokens_per_wall_second and decode_tokens_per_wall_second are aggregate response usage divided by whole-run wall time. They are not server-internal prefill/decode kernel rates. TTFT remains the critical-path usability row and is not hidden by aggregate work. The endpoint emitted no cached-token count, so this run makes no cache-hit claim; #5787 owns the controlled shared-prefix A/B and cache provenance.

Observed overload boundary

The first 16-worker attempt completed only 10/100 contexts before 90 requests exceeded the three-minute header timeout. The artifact is retained at s1-gcp-realendpoint-workers16-overload-2026-08-06.json. This is the first concrete blocker found by S1: software fan-out is cheap, but physical admission must match endpoint capacity. More logical contexts are useful only with bounded workers, backpressure, compatibility scheduling, and overload telemetry—the work tracked by #5788 and #5790.

Contract coverage

• One immutable base is inserted by the endpoint adapter; each microagent supplies only its task delta.
• internal/microagent.Host bounds physical workers and preserves each context’s ID/error.
• Streaming SSE provides observed TTFT and final usage for every successful response.
• The test endpoint verifies the shared base, 1:1 request accounting, cached-token parsing, and telemetry aggregation.
• No terminal, cwd, transcript manager, credentials bundle, or full harness process is created per logical context.