The test suite was reading whoever ran it: v0.6.11, OMNI

Issue #308 added an OMNI_HOME override so a test run would stop touching the developer’s real configuration. It was verified, it worked, and it was checked the wrong way.

The verification measured writes. Nothing wrote to ~/.omni any more, so the override looked complete. But toml_filter::resolve_user_signal_dir derived ~/.omni on its own, which put it outside the override, and it runs on every hooked command as a read.

The consequence is worse than a dirty test directory. Filters from the developer’s home still joined the find() race that decides which signal claims a command, so which filter won depended on whose machine the suite ran on. A test could pass locally and fail on CI for reasons that had nothing to do with the change, and it did: an embedded cargo filter won on the Linux runner and lost on the maintainer’s laptop, which turned a real bug in the pipe’s TOML branch into a failure that looked like a platform problem.

Every path in the tree resolves through one paths module now.

XDG, and the migration decision inside it

While the path resolution was being rewritten anyway, it grew the layered lookup a Unix tool is supposed to have (#217):

config.toml, filters/, signals/, trusted.json
  OMNI_HOME -> OMNI_CONFIG_HOME -> an existing ~/.omni -> $XDG_CONFIG_HOME/omni -> ~/.omni

database, transcripts, caches, exports
  OMNI_HOME -> OMNI_DATA_HOME -> an existing ~/.omni -> $XDG_DATA_HOME/omni -> ~/.omni

Note the third rung. An existing ~/.omni deliberately beats XDG, and that is the migration decision rather than an oversight: an install that already has a tree keeps using it, so upgrading cannot appear to lose your database. Nothing moves on its own. With no environment set, every root is still exactly where it was before this release.

Two more cuts that removed the answer

A grep -n over a single file lost 7 of its 12 matches (#316). grep only prefixes the filename when it was given more than one file, so a single-file grep -n prints <lineno>:<text> and nothing else. The detector required the part before the colon to look like a path, so that shape was not recognised as grep output at all and fell through to a generic fallback that keeps the segments it rates highly. On a 12-match grep over a Grafana alerting manifest the agent received 5 lines: four of runbook prose, one expression. The config lines, which were the reason to run the grep, were the ones dropped.
A pipeline ending in a stage that rewrites its payload is now routed to that stage (#277). The exception added in #269 covered jq and yq only, so kubectl get pods | cut -d' ' -f1 was still handed to the kubectl distiller, which then described a table the command no longer emits.

That second one is worth a note on method. The issue proposed classifying every stage and routing to the last transformer. Measured over 4,958 recorded pipelines, that design is worse at both ends: routing by the last stage regardless hands 69.1% of pipelines to head, tail or grep, all of which are verbatim passthroughs, and stops distillation on them entirely. The measurement rejected the design the issue asked for, so we shipped the narrower exception instead. A measurement that kills your plan is the measurement working.

Dead weight in the hot path

SessionState::scoring_modifier was consulted in two scorer.rs branches, one of them per line, and assigned in no production path (#164). It was permanently None, so both branches were pure cost on every line of every command.

That also settles a question #164 left open. omni goal looked load-bearing because the scorer read something only omni goal could set, and it turned out the scorer could not read it at all. omni goal keeps its place through the SessionStart injection, which is a different mechanism and one that works.

The connection pool

r2d2 defaults min_idle to max_size and build() blocks until that many connections exist, each through the pragma customizer. A hook that exits after one payload was opening four SQLite connections before doing any work (#174).

Median of 10 opens on the release binary: 3.19 ms down to 1.26 ms with min_idle(Some(1)), against 0.81 ms for a bare connection carrying the same pragmas. One line recovers 73% of the cost.

The pool stays. Removing it entirely is 62 call sites for the remaining 0.45 ms, and the MCP server genuinely needs it.

brew install fajarhide/tap/omni && omni init