The benchmark was counting a terminal: v0.6.13

tests/bench_replay.rs selected command, raw_input FROM execution_traces with no filter on which agent produced the row. On the reporting installation that corpus is 68% terminal by bytes: 888 traces carrying 86.0 MB of output a human read in a shell and no model ever received, against 13.9 MB from Claude Code and 25.7 MB from Aider.

Replaying the same pipeline over the same database, one run each:

population	raw	distilled	net
everything	126.0 MB	26.3 MB	79.1%
model-facing only	40.1 MB	22.7 MB	43.3%

The README quoted the first number. It should have quoted the second.

We had already found this bug once. Issue #212 established it in July and fixed it in omni stats, which prints “Terminal output is excluded” on its own line. The harness that produces the number the website quotes never got the same fix. A fix applied to one of two readers of the same table is how a correction gets undone without anyone touching the corrected code.

The harness replays the model-facing population by default now and prints which one it used. OMNI_BENCH_ALL=1 asks for the wider one. Both are worth having, because the gap between them is the mistake, and hiding one of the two is how it comes back.

Every published number re-measured

The README carried 58.9% net over an 1,810-trace corpus with 63.6% of calls saving nothing. Measured now over 9,965 model-facing traces: 43.3% net, 90.0% saving nothing, and not one call in 9,965 that made the output larger.

The drop is not a regression. It is the correctness work of the last few releases showing up in the accounting. Build: ok no longer answers for commands that never built, and make, kubectl listings, single-file grep -n, green npm test and reshaping pipeline tails all pass through now. Each of those passthroughs used to be counted as a saving.

Latency moved the other way. A 496 B git status is 21.1 ms on a fresh database and 60.7 ms on a 205 MB one, against the 82 ms and 308 ms the README promised. A 16.5 KB cargo test is 24.5 ms rather than 276 ms.

The import graph counted directories as modules

resolve_candidate accepted any candidate path that exists(). A brace-grouped use crate::pipeline::{CollapseMode, SegmentationMode} leaves the extractor holding crate::pipeline::, whose candidate is the directory src/pipeline. Directories exist, so they won the lookup and became graph nodes.

Measured on this repository: src held 113 dependents and src/pipeline 30. That is 143 of the graph’s edges pointing at two directories, while src/pipeline/mod.rs, the file all thirty of those imports actually name, had none. omni_context reported that number.

One word fixes it, is_file() for exists(), because mod.rs and index.* were already later in the candidate list.

The graph will not be wired into the scorer

The proposal was to feed imported_by.len() into scorer::score_segments, so a file forty others import survives a cut that a leaf file does not. It was sampled against 2,764 recorded traces before any of it was built.

Only 127 name a file this repository’s graph knows. 52 name two or more. 26, under 1%, name two files whose dependent counts differ, which is the entire population a dependents signal could re-rank. Building the persistence and mtime-invalidation layer the issue calls “the actual work” buys a tie-break on fewer than one call in a hundred, against a build_graph that takes 542 ms where the budget for the whole hook is 10 ms.

The sampling also found the graph was wrong, which is the fix above. Before it, the same probe read 196, 95 and 47, because two directory nodes were absorbing the edges.

A grep at the end of a pipeline owns its output

kubectl … logs … | grep -iE 'mcp|slack|error|warn|ready' delivered one line of fifteen. Routing reads a pipeline as belonging to its first stage unless the last one reshapes the payload, and grep reshapes nothing, so kubectl kept the buffer and its distiller kept is_critical lines only.

The line it kept was the ERROR. The fourteen it dropped included the WARNING the pattern named by name, 3/3 MCP servers connected and Bolt app is running!. What arrived said the pod had failed to start. The evidence that it had started was behind the retrieval marker.

A grep pattern is the caller’s filter. That rule now sits at the routing boundary: a grep, rg or ag tail claims the payload, and the grep path is the only distiller allowed to touch it.

brew install fajarhide/tap/omni && omni init