Format-safe compression, and honest numbers: OMNI v0.6.2

v0.6.2 is the release where the project’s numbers changed direction. Everything below made the advertised saving smaller, and that is the point of it.

Structured output was being corrupted

The pipeline had no idea what kind of text it was compressing. Collapse squashed the repeated lines of structured output into [N similar lines collapsed] and handed back a payload that no longer parsed. A JSON dashboard piped through the hook came out carrying 14 collapse markers and failed jq outright, which means every downstream jq, json.load or kubectl apply in the agent’s plan failed with it.

A new format sniffer (pipeline::format) now gates both choke points, hooks::post_tool and hooks::pipe. JSON, NDJSON, YAML, TSV and CSV pass through byte for byte at zero marker cost. Plain text still compresses exactly as before.

A YAML file that came back as a docker log

sniff_yaml judged every line of a document that had no leading ---. A ConfigMap carrying Vault HCL under config.hcl: | therefore looked like free text, because 26 of the manifest’s 608 lines were HCL rather than YAML. The format gate stood down, collapse emitted [8 similar lines collapsed] markers, and then is_docker_logs counted those markers as timestamp prefixes: 12 triggers against a threshold of 5.

The result: 13,463 bytes of Kubernetes config came back as docker logs: 323 lines, no errors detected. OMNI’s own collapse markers fabricated the evidence that misled the next stage. A block scalar’s body is now skipped rather than judged, because the | that opened it is YAML’s own statement that what follows is a value.

find and grep stopped deleting their payloads

The v0.6.1 find behaviour, credited with “up to 99.8% savings”, reported about 99% by discarding the file paths that were the answer. It is replaced with lossless prefix factoring: hoist the shared directory prefix into a header, state each path once relative to it, cut at a separator so every path round-trips byte for byte. Real saving on the same output is about 72%, with nothing lost.

grep now hoists each path once instead of repeating it on every match, which is 26% to 39% and also lossless. Both hand the input back untouched when factoring would make it larger.

A weak filter was hiding the test distiller

sys_build_domain matches every cargo invocation but strips only Compiling-style lines. It won the find() race, shadowed TestDistiller, cut 1%, and had that 1% thrown away by the MIN_REDUCTION_PCT guardrail. So cargo test reported 0% on output the real distiller reduces by 94%. A TOML filter now only short-circuits the distiller if it actually beat the guardrail. A filter that earns its match still wins, user filters included.

With the distiller reachable again, a second bug surfaced: it counted result-line segments and reported 1 passed for a run cargo itself called 490 passed, because CollapseMode::Test folds the 330 ... ok lines away before the distiller ever sees them. It now quotes the runner’s own summary line (cargo, pytest, jest) and only falls back to counting when no summary was printed.

Every published number, re-measured

The headline claims were not stale. They were untrue.

• docs/PERFOMANCE.md sold docker build as 9.2 KB to 49 bytes (99.5%). The fixture it names distils to 5,783 bytes, which is 37.2%.
• git diff quoted 50.0% on bytes (397 to 220) that compute to 44.6%.
• The 97.3% all-time figure traced to find . at minus 12.6 million tokens, the defect fixed above. The number was manufactured by a bug.
• $35+ USD/month priced a cost estimator that does not work (see below).
• ~40% faster TTFT was never measured.
• Three testimonials in quote marks were attributed to nobody, because nobody said them.

They are replaced with a replay of 1,810 real execution traces on the release binary with a fresh HOME per invocation: 58.9% net, 15.0 MB down to 6.2 MB. Per command: git 91.3%, cargo 96.8%, cat 9.1%. Fixing the YAML sniffer moved the reported saving down, from 65.3% to 58.9%, because six of those points were destroyed manifests rather than removed noise. All six translated READMEs carry the same figures.

The dollar estimator is gone

est_cost_usd and ModelPricing are removed from stats, diff and guard::config. OMNI is a hook. It never sees the API’s usage block, so it cannot know whether the bytes it saved were billed as fresh input or as a roughly 10x cheaper prompt-cache read. ModelPricing used only .input, its _cached and _cache_creation fields were dead, and the formula was duplicated across four call sites. Every dollar figure OMNI printed was a fresh-input guess presented as a session cost. Unknown keys in an existing config.toml still parse, so nothing breaks on upgrade.

Two costs we now document instead of hiding

OMNI is not deterministic against a warm database. Session history feeds the scorer, so the same binary on the same input differed on 21 of 30 traces run to run. One gave 1,835 bytes, then 433. Any A/B measurement has to isolate state per invocation.

Latency is real and it grows. Roughly 82 ms for a 496-byte git status against a fresh database, and roughly 308 ms against a 97 MB one, per hooked command.

One more thing: 0.6.2 almost did not ship

release.yml asked dtolnay/rust-toolchain@stable for each matrix target, which installs that std into stable. But rust-toolchain.toml, added this release, makes cargo switch to the pinned 1.97.0, which has std for no target except the runner’s own. Every cross-compile died with error[E0463]: can't find crate for 'core' before compiling a line of OMNI, and fail-fast took the rest of the matrix down with it, so the tag produced no binaries at all. The matrix targets are now listed in rust-toolchain.toml itself. ci.yml stayed green throughout, because it only builds host-native, which is why this was invisible until a tag was pushed.