seek-router

A PreToolUse hook that turns an agent's grep / rg / git grep calls into seek searches, plus the seek-search skill that teaches the query syntax.

Install and usage: see the Agent Integration section of the top-level README.

How it decides

bin/router.sh reads one hook payload on stdin and prints at most one decision on stdout. There are exactly two outcomes:

Outcome	When
`allow` + `updatedInput`	the whole command is a plain file search it can translate
nothing at all	everything else — the original command runs untouched

It rewrites rather than denies. A denial makes the model produce a replacement command, and a hook that keeps rejecting the same command can put the model in a retry loop; rewriting sidesteps that, and the tool result is not marked as an error.

The notice rides along in additionalContext, so the rewritten command stays a single seek invocation with one interpolated value. That value is shell-quoted: a pattern containing a quote must not be able to close ours and leave the rest running as shell.

Always exit 0

Claude Code reads exit code 2 as a deny and shows the hook's stderr to the model as the reason. A crashing router would therefore look like a deliberate block and break search instead of falling back to grep. Every path exits 0, including malformed input, and make test-plugin asserts it on every fixture.

Layout

.claude-plugin/plugin.json   Claude Code + Codex manifest
plugin.json                  Agent Plugins 1.0.0 manifest (skill only)
hooks/hooks.json             matchers and if-conditions
bin/router.sh                the router
skills/seek-search/SKILL.md  query syntax, ships everywhere
test/run.sh                  fixture suite, run by make test-plugin

Not built: routing the Grep and Glob tools

Claude Code also exposes Grep and Glob tools, which this plugin leaves alone. updatedInput can change a tool's arguments but not which tool runs, so routing those means denying the call and returning seek's output in the deny reason. That works — the model reads results out of a denial and answers from them — but it costs considerably more than the Bash path:

a denial can loop. Nothing caps consecutive PreToolUse blocks the way the Stop hook's limit does, so the router would need its own circuit breaker keyed per session and subagent.
seek would run inside the hook rather than in the agent's shell, which puts index-build time on the hook's timeout instead of in plain sight.
Grep's output_mode, head_limit and context flags would each need a faithful mapping, and several have none — count in particular.

Whether it is worth building depends on how often the model reaches for the Grep tool rather than shell grep in normal permission mode, and on whether routing improves an answer rather than merely producing one. Judging that needs end-to-end tasks scored on output quality, not a token count.