The escape hatch the hook never read: v0.7.4

A release about claims that were not true. A documented escape hatch that did nothing on the host it mattered on, a setup command an agent could not run, a redaction rule that ate valid code, and a configuration layer worth 0.031% of the corpus it slowed down.

The escape hatch was inert on the host that carries the traffic

The manual says OMNI_PASSTHROUGH=1 <your command> skips the pipeline entirely, raw output every time. It appears on four pages, including OMNI_PASSTHROUGH=1 kubectl get pods -o yaml, which is exactly the case where somebody needs the exact bytes.

It was true for omni exec and for the pipe, where one process both reads the variable and does the work. It was false on the Claude Code hook path. The PostToolUse hook is a process the host spawns, so it inherits the host’s environment and never sees an assignment written in front of a command. The prefix did nothing, and nothing said so.

That is the worst shape a bug can take here: the hatch works everywhere you test it by hand, because every hand test uses omni exec. The hook now parses the leading KEY=value off the command string it is handed, so the documented form behaves as documented.

Setup needed a terminal, and the audience is agents

omni init with no flags opened a dialoguer menu, which can only fail without a tty. It exited 1 on IO error: not a terminal and named no remedy. That is the line the README, the Homebrew caption, this site and install.sh all print, so the one documented instruction was the one an agent could not follow.

With no terminal it now configures the host that ran the command, read from the environment the same way the exec and pipe paths already read it, and says which host it picked. A host it cannot name, a plain shell included, gets an error listing the flags rather than a guess. Falling back to --all was considered and rejected: fourteen host configs and a file written into the working directory is too large a side effect to trigger because stdin happened to be a pipe.

There is also a plugin, which makes the whole thing two lines typed inside the session:

/plugin marketplace add fajarhide/omni
/plugin install omni@omni

It installs a skill rather than the binary, and that is the gap the MCP server could never close. An agent with omni_retrieve wired still had nothing telling it when to call it, or what to do when output arrives carrying a marker.

Redaction was deleting code and calling it a credential

key={item} in a .tsx file came back as key=[REDACTED]. key is the prop React requires on every list item, and a formatter puts it on its own line as soon as the element has a few attributes, so it is the most common key= line in a component. The redacted form is not valid TSX, so an agent that reads a file to edit it and writes it back produces a syntax error.

The same rule was found eating this project’s own test suite: PASS=$((PASS + 1)) in tests/smoke_test.sh arrived as PASS=[REDACTED].

A brace expression is code, and a shell expansion names a value instead of holding one. Both are delivered as written now. API_KEY=sk-ant-... and PASSWORD="$3cr3t!" are still cut, and the second one is the reason the rule checks what follows the $ rather than the character alone.

The filter layer is gone, and the savings did not move

Re-measured on the same 6,656-trace corpus with the loader stubbed to empty: the aggregate is 2.7% / 14.9% with the TOML layer and 2.7% / 14.9% without it. The whole layer was worth 2,018 bytes over 6,656 commands, 0.031% of the corpus, and infra output scored slightly better without it.

Priced by removal on the release binary rather than by a unit-test timer, it cost 5 to 7 ms on every hook against a 10 ms budget.

	before	after
aggregate savings	14.9%	14.9%
p90 per hook	21.4 ms	10.5 ms
dependency tree	265 crates	237
release binary	9.93 MB	9.17 MB

What replaces it is nothing. A tool that needs handling gets a Rust distiller, which is snapshot-tested and cannot be shadowed by whichever regex matched first. clap went the same way: it was 141 lines of derive deciding seven boolean flags, in front of eighteen modules that parsed argv a second time anyway.

The manual is now checked against the code

tests/docs_match_the_code.rs asserts that every omni <cmd> written inside a shell fence in these pages is a subcommand the binary actually has, and it reads the list from main.rs rather than keeping a copy. A page that documents a command which exits 1 now fails the build instead of the reader.

Full notes are in the 0.7.4 changelog.

brew upgrade omni     # or: omni update