Output shaping

Input compression is the well-known half of the cost problem and the smaller one. Output is billed at several times the input rate on reasoning-capable models, and a large share of what is generated is preamble, restated context and narration of work nobody asked to see. A turn that merely resumes after a successful tool result rarely needs full reasoning effort, yet most gateways send the same setting for every call in a run.

Two levers, both clamp-only.

from tesserix_adk.models import Effort, Shaping

shaping = Shaping(
    enabled=True,
    baseline=Effort.HIGH,
    resumption=Effort.LOW,
    terseness="Answer directly. No preamble.",
)

Both are off until an agent turns them on.

Clamping effort

shaped = shaping.shape(request, effort=Effort.HIGH, final=is_last_turn)
body |= provider_effort(shaped.effort, provider=provider.provider_name)

shape reads the request and returns a decision; it never modifies the request, so the cacheable prefix is exactly what the assembler produced. shaped.effort is never above what the caller asked for — the clamp is a min, and a policy whose resumption exceeds its baseline raises ConfigurationError at construction rather than at call time.

What is not clamped, and why:

Turn	Effort	Why
New question	as asked	the deliberation is the answer
Tool resumption	clamped to resumption	the work was done by the tool
Error recovery	as asked	failures are where deliberation pays
Final turn (final=True)	as asked	the saving is small; what it degrades is what the user reads
Structured output	as asked	the schema requires completeness

A turn is classified from what the conversation ends with. Mark a failed tool result with errored so the recovery is not clamped:

messages.append(errored(result) if outcome.failed else result)

provider_effort maps the one expression of effort onto each provider's own parameter, and returns nothing for a provider that has none — a local llama.cpp build has no effort setting, and that is not an error.

shaped.attributes() gives the model call record what was applied and what was asked for. Settings and outcomes only; no prompt content.

Steering terseness

request = shaping.steer(request)

The instruction is appended after everything already in the system prompt. Nothing existing is rewritten, because a rewrite invalidates the cached prefix and gives back more than the shorter output saves. The suffix is byte-stable and applied idempotently, so enabling it costs exactly one prefix invalidation and none afterwards. Nothing below the system prompt is touched; where there is no system message, one is added.

Known limitations

• Classification reads the tail of the conversation only. An agent that interleaves its own scratch messages after a tool result should call shape with the turn it knows it is on.
• The effort vocabulary is the kit's four levels. A provider offering a token budget rather than a level is mapped at its own adapter; provider_effort returns nothing for it here.
• Steering alone measures nothing. Use savings accounting with a stable control holdout to label input savings as measured and output savings as estimated.