harness-design

Overview ¶

Command harness-design is a development rig, not a shipped surface.

THE QUESTION IT ASKS: handed a plain goal and nothing else, can a MODEL architect a good sub-harness — and does the harness it wrote produce a good answer when it is actually run?

That question is why nothing here is hand-written. A fixture somebody typed proves the runner works, which internal/subharness already proves in its own tests; what it cannot prove is that the DESIGN half is reachable by a model from words a person would say. So the rig has four stages and no shortcuts between them:

STAGE 1    DESIGN  the meta-guide (internal/subharness/prompts/designer.md,
                   rendered by designer.go) + the goal → a harness page, its
                   cues and its justification. Salvaged, decoded and validated
                   by internal/subharness, repaired once, then retried with
                   the error fed back.
STAGE 1.5  REVIEW  the same guide wearing PART FOUR reads the draft as a
                   critic — SPEED, COST, QUALITY — and hands back a PATCH.
                   The ops are applied to the draft this rig already parsed,
                   so the critic never retypes what it is not changing, and
                   the delta printed is the patch itself.
STAGE 2    PRINT   the page, the justification, and the CARD — the rendering
                   a person actually approves.
STAGE 3    RUN     the page executed against a model-backed Env
                   (execmodel.go), with the full trace and every step's whole
                   output printed, and the trace saved under
                   harnesses/<name>/run/<ts>.json.

AND A SECOND RIG BEHIND THE SAME BINARY. `-orchestrate` runs none of the above: it asks the opposite question — what if nobody designs a graph at all, and a planner amends a live frontier on every completion? That is internal/orchestrate, and orchestrate.go is the whole of it. The two share this file's transport, salvage ladder and printing, and nothing else.

Usage:

OPENROUTER_API_KEY=… go run ./cmd/harness-design -goal all
OPENROUTER_API_KEY=… go run ./cmd/harness-design -goal hard
OPENROUTER_API_KEY=… go run ./cmd/harness-design -goal "your own sentence"
                     -model deepseek/deepseek-v4-flash -design-only -review=false
OPENROUTER_API_KEY=… go run ./cmd/harness-design -orchestrate -goal all -fuel 1.50
OPENROUTER_API_KEY=… go run ./cmd/harness-design -repro 5 -belt chat
                     -model deepseek/deepseek-v4-pro -goal "…"

AND A THIRD, WHICH IS AN INSTRUMENT RATHER THAN A QUESTION. `-repro N` runs stage 1 N times and tallies what refused it — see repro.go, and the reason it had to exist.