Local context efficiency for coding agents

Smaller tool results.
Exact recovery.

Laconic transforms noisy file, command, and search results before they enter an agent's context. The original stays in a private local ledger, available by exact full or ranged expansion.

See the transformation How it works

Strict-smaller decision

3,682 → 286 characterspages-demo/F1Full or ranged exact expansion

Deterministic transformation

A real codec run.
Every omitted byte recoverable.

This committed fixture passes through Laconic's file encoder and exact recovery envelope. The complete envelope is strictly smaller than the original before it is shown.

READ · 3,682 raw characters

[laconic pages-demo/F1 | full: laconic_expand({"reference":"pages-demo/F1"})]
demo/recovery.py  108 lines
  outline: RecoveryEntry:9-11  checkpoint_1:13-19  checkpoint_2:21-27  checkpoint_3:29-35  checkpoint_4:37-43  checkpoint_5:45-51  checkpoint_6:53-59  checkpoint_7:61-67  [+5 more]

286 visible characters · 92.23% fewer
Original retained locally under pages-demo/F1.

Exact recovery proof

Full and ranged.

Full handle: pages-demo/F1
Range handle: pages-demo/F1:1-12
Recovered SHA-256: 486976ae75cac4eb75dd118e638f9d9c8307699fe757a4ba1420ec98d37df5cc

"""A deterministic recovery ledger used by the public Laconic demo."""

from __future__ import annotations

from dataclasses import dataclass


@dataclass(frozen=True)
class RecoveryEntry:
    key: str
    value: str

Observed

While developing Laconic.

Actual codec counters, separate from economic modelling.

Fewer visible chars: 6,900,697
Boundary reduction: 28.86%
Exact expansions: 117

Frozen local Laconic-repository cohort only. Not representative of other users, repositories, hosts, or workloads. Inspect the evidence.

The system

Reduce first.
Recover exactly.

The host adapter sends a completed tool result to one canonical runtime. The runtime accepts a recovery-bearing envelope only when exact expansion succeeds and the complete visible result is strictly smaller.

Five-step Laconic path from host tool result through runtime, encoder, private ledger, and recovery-bearing envelope.

Fail-open contract

Compression is optional.
Correctness is not.

Every decision preserves the host's result unless a complete, exactly recoverable, smaller replacement has already been proven.

Strictly smaller

Laconic counts the complete recovery-bearing envelope, not an optimistic intermediate encoding.

Raw retained first

The original remains in the private local ledger until recovery and size checks succeed.

Original on failure

Unsupported tools, errors, malformed shapes, and non-smaller candidates return the original unchanged.

Install entry point

Use the maintained instructions.

The repository README owns installation and host setup. The site does not fork a second guide that can drift.

Open installation instructions

Claim boundary

What happened.
What is modelled.

Observed character counters and modelled economic implications never share a label, card, or color. Unsupported token, cache, behavior, and general-benefit claims remain out of scope.

Observed

6,900,697

fewer characters visible at the tool-result boundary

Raw → visible: 23,914,138 → 17,013,441
Emitted: 1,225

modelled_not_measured

$33.89–$76.00

3.87–8.69% of the named modelled denominator

Modelled denominator: $874.73 across 85 joined sessions
Host-reported partial denominator: $887.24 across 85 reporting sessions
Fallback-priced share: 0.000000%

Single-arm model with no counterfactual. Not measured, billed, token, cache, or behavior savings.

Read the frozen evidence ledger and its methodology and limitations.