Skip to the content.

Micro-context S4a: harness assumptions and lightweight descriptor

Status: observed local contract/bring-up witness, 2026-08-06. This is the harness-assumption child #5789, not the tool/effect-safety S4 child #5791.

fak-micro-context-descriptor/1 captures only the semantics needed to schedule one logical context:

{
  "schema": "fak-micro-context-descriptor/1",
  "id": "d-0001",
  "base_id": "immutable-agent-base-v1",
  "task_delta": "task-0001",
  "capability_set": ["read_record"],
  "budget": {"max_turns": 1, "max_output_tokens": 8},
  "continuation": [],
  "output_contract": {"kind": "exact", "expected": "DONE"}
}

The adapter uses the existing microagent.Host and Gateway: it expands a registered immutable base plus continuation and task delta into one model turn, advertises only the named tool capabilities, applies the output-token budget, and refuses an output outside the contract. It does not create a second runtime.

Assumption inventory

Assumption Classification Carrier / disposition
model endpoint and shared instructions semantic base_id, resolved once by adapter
task input semantic task_delta
tool authority semantic explicit capability_set
bounded work semantic budget (v1 is exactly one turn)
prior state semantic continuation
accepted result semantic output_contract
one OS process per context session convenience omitted
terminal/TUI per context session convenience omitted
cwd adapter concern capability/resource-lease adapter
ambient credentials unsafe convenience omitted; explicit capability adapter
interactive approval UI adapter concern policy/effect adapter

This is where Ultracode-style decomposition helps: independently specified deltas and output contracts are useful. Its concurrent orchestration is prior art only; it is not inference/cache evidence and is not embedded as another scheduler.

Reproduce

go run ./cmd/microcontextdemo `
  -descriptor-bench experiments/microcontext/s4-local-descriptor-benchmark-2026-08-06.json `
  -contexts 1000
go run ./cmd/microcontextdemo `
  -verify-descriptor-bench experiments/microcontext/s4-local-descriptor-benchmark-2026-08-06.json

Observed envelope

The checked-in run executed 1,000 descriptors through one existing 16-worker microagent.Host and one deterministic gateway fixture:

The CLI numbers are deliberately narrow process/version probes. They do not measure full harness task completion, child-process RSS, prompt tokens, model quality, or Ultracode orchestration. The in-process fixture is likewise not model TTFT or throughput. A fair full-task corpus and per-process RSS still require harness-specific noninteractive adapters; the current witness names that limitation rather than manufacturing a ratio.