JevTO: output selection for coding agentsExperimental · v0.1.1Experimental

Less tool output.
Exact recall.

Give Claude Code and other coding agents a smaller view of command output. JevTO selects real lines, preserves exit codes, and stores the original bytes locally for recall. Start with rules only; no API key needed.

Command output: −94%
Required facts visible: 45/46
Session input, Codex: −9%

Install JevTO Watch the demo

Free · Apache-2.0 · experimental v0.1.1. Payload reduction is not measured bill savings. Read the evidence.

Output inspector

Same command. Three views.

REAL BENCHMARK CAPTURES

exit 1

Agent's goal: Fix the failing parser test

Loading capture…

Exact text each tool's Claude Code hook delivered in the benchmark run. Temp paths shortened to <tmp>.

Deterministic mode: no network, no accountRecall returns historical bytes; it never reruns a command

How JevTO selects outputMethod

Keep a useful view. Recall the rest.

Passing-test inventories, progress redraws, and regenerated lockfiles can fill an agent's context. JevTO selects output line by line and records why it defers a section. Selection can miss useful evidence, so every view includes a way to recover the captured bytes.

1
Capture exactly
The command runs with your permissions. Stdout, stderr, and the exit code are stored separately and hashed in a local store.
2
Protect recognized failures
Recognized failures, assertions, panics, warnings, and exit status are protected by local rules. Jev ranking cannot override those protections.
3
Fold the ceremony
Recognizers for Rust, Go, Python, Node, Jest/Vitest, and TAP test inventories, Cargo progress, lockfile and minified diffs, and long logs replace noise with a count and a gap marker.
4
Recall on demand
Each gap names a section. jevto recall ID --section … returns those bytes exactly. If the view would not be smaller, the original is delivered unchanged.

What the command printed

test_accepts_equals ... ok
test_accepts_tabs ... ok
… 218 more ok lines …
test_rejects_leading_space ... FAIL
AssertionError: '7' is not None
Ran 221 tests in 0.021s
FAILED (failures=1)

What the agent reads

jevto exit=1 omitted=225
... 220 passing tests hidden [stderr-0-13420]
test_rejects_leading_space ... FAIL
AssertionError: '7' is not None
Ran 221 tests in 0.014s
FAILED (failures=1)
recall: jevto recall d3768e28

3,572 → 200 tokens on this run. The gap marker keeps the count, so the agent still knows 220 tests passed.

Optional: Jev, the decision layer.

Jev can rank search results, diffs, and long-output sections by meaning. The default auto mode enables this when an OpenRouter key, a goal, and eligible output are present. Bounded snippets or line-shape digests and the goal may be sent to OpenRouter (≤ 64 KiB, 8 s timeout, secret filtering). Use JEVTO_MODE=rules for no network calls. The historical live payload run kept 46/46 benchmark facts visible; rules kept 45/46. Understand what leaves your machine.

Capture → Protect → Fold → RecallEvery decision is written to a local receipt

BenchmarksReproducible

Fewer tokens only count if the answer survives.

23 generated scenarios cover tuned dev cases, a holdout, and vocabulary-mismatch cases. Each tool uses its own hook where routed; explicit-wrapper cases are labeled in the raw report. Tokens are bytes ÷ 4 for one command's output. A fact is a required string visible without recall, such as the failing test or relevant commit. The final holdout results include fixes made after its first run.

Historical rules-only payload totals by suite. Native: 269,736 estimated tokens, 46/46 facts. RTK 0.48.0: 236,535, 44/46. JevTO rules: 16,979, 45/46.

Log scale. Hollow red-marked dots lost a required fact. RTK's hook doesn't route Python, Node, or plain log dumps, so those rows sit at native.

A 98% smaller test result is a much smaller session win.

Same task on two agents: implement a small parser, verify with a 260-test unittest -v suite, 15 hidden holdout cases. Each run starts from a fresh copy of the repository, three runs per arm. The command output shrinks 50×, but the system prompt, file reads, and the agent's own reasoning still dominate the bill.

Left: test output read per run, about 17.7 KB native versus about 350 bytes with JevTO. Right: session tokens per run for each arm on Claude Code and Codex, showing overlapping ranges with JevTO lowest at the median.

Codex CLI 0.158 · gpt-6-sol

Median of 3	Input	Uncached	Time
Native	198,520	25,969	69.5 s
Ponytail	214,289	28,695	86.7 s
JevTO	180,533 −9%	17,973 −31%	64.6 s
Both	222,212	23,153	72.2 s

12/12 sessions passed the visible suite and hidden holdout.

Claude Code 2.1.283 · Haiku 4.5

Median of 3	Tokens	Est. cost	Code lines
Native	120,494	$0.060	30
Ponytail	125,662	$0.065	27
JevTO	116,584 −3%	$0.045 −25%	31
Both	107,819 −11%	$0.051	24

All 12 passed visible tests and failed the same strict holdout. These are failed task outcomes, not verified successful completions. Cost is a list-price estimate, not a bill.

Historical rules-only payload comparison · estimated tokens = bytes ÷ 4 · facts visible without recall
Scenario	Native	RTK 0.48.0	JevTO rules
Dev set · rules tuned here
rust-test-failure	1,245 · 3/3	126 · 3/3	155 · 3/3
go-test-failure	2,651 · 2/2	64 · 2/2	78 · 2/2
python-unittest-failure	3,572 · 3/3	3,572 · 3/3	200 · 3/3
python-quiet-warning	3,397 · 2/2	3,397 · 2/2	89 · 2/2
node-test-failure	1,576 · 2/2	1,574 · 2/2	197 · 2/2
rust-test-lean	119 · 1/1	10 · 1/1	61 · 1/1
service-log-triage	52,163 · 2/2	52,163 · 2/2	2,075 · 2/2
search-many-hits	3,688 · 1/1	3,688 · 1/1	356 · 1/1
git-diff-lockfile	11,484 · 2/2	766 · 2/2	120 · 2/2
git-diff-multi-file	7,225 · 1/1	5,738 · 0/1	1,109 · 1/1
git-log-history	12,618 · 1/1	222 · 0/1	551 · 1/1
Holdout · written after the initial freeze
cargo-multi-failure	4,311 · 6/6	205 · 6/6	280 · 6/6
tsc-type-errors	70 · 3/3	90 · 3/3	70 · 3/3
go-package-failure	129 · 2/2	58 · 2/2	99 · 2/2
node-all-pass	2,571 · 2/2	2,574 · 2/2	48 · 2/2
python-stdlib-traceback	1,261 · 3/3	1,261 · 3/3	389 · 3/3
rg-todo-sweep	1,913 · 1/1	1,913 · 1/1	215 · 1/1
git-diff-rename-plus-fix	2,233 · 2/2	1,849 · 2/2	835 · 2/2
cargo-build-warnings	2,084 · 3/3	2,026 · 3/3	2,084 · 3/3
json-log-triage	108,527 · 1/1	108,527 · 1/1	3,351 · 1/1
Semantic · vocabulary-mismatch cases
semantic-log-signout	42,783 · 1/1	42,783 · 1/1	1,411 · 1/1
semantic-search-deadline	1,132 · 1/1	1,132 · 1/1	222 · 0/1
semantic-diff-rounding	2,984 · 1/1	2,797 · 1/1	2,984 · 1/1
Total, 23 scenarios	269,736 · 46/46	236,535 · 44/46	16,979 · 45/46

Historical results: 2026-09-30 on Windows, JevTO 0.1.0 and RTK 0.48.0. These are not a new v0.1.1 benchmark or current-competitor claims. Method, limitations, and reproduction · Raw views and JSON.

honest-reading.txtLimits

What the numbers do and don't say.

Where it wins

Test runners RTK doesn't route (Python, Node) and long logs: −69% to −98% with every fact kept, on scenarios written after the rules froze too.
Lockfile-heavy diffs: 11,484 → 120 tokens; a 14-file mechanical refactor shows the edit once and the real fix in full.
Keeps facts that truncation drops: the git log commit and the multi-file fix line.
The small Codex pilot used fewer median input tokens with 12/12 hidden-holdout passes across its four arms.

Where it doesn't

RTK is tighter on Rust and Go failures and tiny passing runs.
Rules missed one vocabulary-mismatch search answer. The historical live Jev arm recovered it, but that does not establish universal answer preservation.
The Claude pilot failed every hidden holdout. Smaller payloads and list-price estimates do not establish lower bills or successful tasks.
n = 3 on one small task per agent. A Cursor pair had unequal outcomes. Indicative, not significant.

The benchmark scenarios were written by the JevTO author. They're generated from scratch by a committed script so anyone can rerun them, or add scenarios that JevTO handles badly.

Supported agentsProperties

The route is explicit.

Every installer previews first, applies only with --apply, backs up what it changes, and removes only its own entry. A failed hook leaves the command native. The Claude hook never sets a permission decision.

Host	Route	Status
Claude Code	`init-claude-auto`: rewrites simple test, build, lint, search, and `git diff/log/show` commands; records the prompt as the session goal	Live sessions
Codex	`init-codex-pre-hook`: PowerShell commands on the same allowlist (`cargo`, `rg -n -H`, `python -m unittest`, …), Windows only	Live sessions
Cursor	Project rule for one exact verifier; opt-in exact-command MCP runner	Narrow
Any MCP host	Read-only `jevto_recall`, `jevto_review`, `jevto_status`	Implemented
Any shell	`jevto run -- PROGRAM ARGS…`	Implemented

Setup, tested versions, and removal instructions. jevto doctor --json reports coverage and mode without making a network call.

Local storagePrivacy

Deterministic mode makes no network requests and sends no telemetry.
Captures expire after 24 hours; the store is capped at 1 GiB.
Command arguments and environment variables aren't stored.
In default auto mode, an OpenRouter key enables eligible remote ranking from the current workspace. Set JEVTO_MODE=rules to stay local.

The store isn't encrypted at rest, and captured output can contain secrets. Read storage, network, and removal details.

Command PromptRust stable

One binary. Start with one command.

Download v0.1.1 for Windows x86_64, Linux x86_64, or macOS Intel/Apple silicon. Every archive has a checksum. No JevTO account is needed.

Prefer source? Use current stable Rust and the pinned command alongside. Installation and first run · Claude Code setup.

C:\src\jevto>

cargo install --git https://github.com/mattbaconz/jevto --tag v0.1.1 --locked jevto
jevto doctor

# Claude Code: preview, then apply
jevto init-claude-auto --workspace .
jevto init-claude-auto --workspace . --apply

# Any agent or shell
jevto run --mode rules -- cargo test --workspace
jevto recall last
jevto gain

Less tool output.Exact recall.