When you ask an LLM for structured JSON (via OpenAI response_format json_schema, OpenAI tools[], or Anthropic tools[].input_schema), you pay an invisible output token tax on every single record.
Ask for 40 invoice line items with 7 fields, and the model repeats those 7 JSON key names 40 times. That is 280 key emissions spent on metadata the receiver already knows from the schema.
As a Technical Integrations Specialist working with REST APIs and webhooks, I wanted to fix this at the wire level—so an application can change a single environment variable (OPENAI_BASE_URL or ANTHROPIC_BASE_URL) without rewriting prompts or application models.
That became wiretuple (MIT, Go 1.23, zero external dependencies).
The lossless positional bijection
Given a deterministic key ordering (required fields in schema order, followed by remaining properties sorted lexicographically), an array of objects:
[
{"line_identifier": "li_0000", "product_sku": "SKU-00000", "quantity_units": 1},
{"line_identifier": "li_0001", "product_sku": "SKU-00001", "quantity_units": 2}
]
is mathematically bijective with an array of positional tuples:
[
["li_0000", "SKU-00000", 1],
["li_0001", "SKU-00001", 2]
]
provided the request schema carries one legend (Tuple: [line_identifier, product_sku, quantity_units]) and closes the array with JSON Schema Draft 2020-12 prefixItems and items: false.
Four pieces in pure Go stdlib
-
Break-Even Gate (
internal/gate): A legend costs input tokens once. Before rewriting a schema, the gate compares expected output savings against the fixed legend cost. Small single-instance payloads stay byte-identical at zero overhead. -
Bidirectional Schema Compiler (
internal/compiler): Preserves unknown JSON Schema properties viaExtraround-tripping and handles nested object tuples recursively (KindTuple). -
Incremental Byte-Level SSE Rehydrator (
internal/rehydrator): No regex and no full-stream buffering. A bounded byte-level state machine handles multi-byte UTF-8 characters split across SSE frames, escaped quotes (\"), brackets/commas inside string literals ("[1,2]"), multi-tool call routing (requestPlan), and falls back to raw output if a model violates arity—keeping the SSE connection alive and JSON valid. -
Passive Audit Mode (
--shadow,internal/shadow): Leaves live traffic untouched while simulating savings into a pure-Go SQLite 3 writer (shadow_records: timestamps, route, token counts, percentages, latency, activation flag—zero prompt or completion payloads ever stored).
Deterministic benchmarks (make bench)
Measured with make bench using a deterministic 4 UTF-8 bytes/token size model and verified with reflect.DeepEqual plus legend key order assertions:
| Scenario | tokens_before | tokens_after | output_savings_% | net_savings_% | p99_overhead_ms | equality |
|---|---|---|---|---|---|---|
| E-commerce invoice line items (40x7) | 2041 | 801 | 60.8% | 59.2% | 0.095 | OK |
| MCP tool invocations batch (25x7) | 1617 | 711 | 56.0% | 53.7% | 0.084 | OK |
| Dev-tool pull request batch (30x7) | 1602 | 889 | 44.5% | 43.0% | 0.078 | OK |
Try it safely first with --shadow
go build -trimpath -ldflags="-s -w" -o bin/wiretuple ./cmd/wiretuple
export OPENAI_BASE_URL=http://127.0.0.1:8787/v1
./bin/wiretuple -upstream https://api.openai.com -shadow audit.db
Source and static release builds: github.com/Dimoncelli/wiretuple

Top comments (0)