feat(hook): wire TOML filters + better tee tail hint

[aeppling]

Summary

Wire TOML-covered commands into the transparent hook rewrite path, and make TOML truncation reversible via the shared tee hints.

Related issues & pr

ISSUES

#2179
#820
#2721

PRs

#2226
#894
#2722
#1068

Part 1: Hook wiring (#2179)

The hook only consulted the static RULES table, so commands covered solely by a TOML filter (jj, jq, ssh, …) weren't rewritten. rewrite_segment_inner now falls back to the TOML registry and rewrites them to rtk <cmd> (guarded against denylisted/reserved names). Builds on #2226.

Part 2: Reversibility

Truncated TOML output now carries a recovery hint: the line-offset tail hint ([see remaining: tail -n +N …]) for contiguous drops, the full-output hint otherwise, and never an unrecoverable marker (falls back to full raw when tee is off).

Then with the command hint

Part 3: Trust guards

Both the project .rtk/filters.toml and the user-global ~/.config/rtk/filters.toml are now trust-gated (SHA-256). rtk init prompts to enable detected filters, or --trust-filters / --no-trust-filters for scripts.

Note: the registry load now hits the hook's Unsupported path (~3–8 ms, one-time per process; the Supported path is unaffected) , a match-only RegexSet is a documented follow-up.

"rtk init -g" (--trust-filters or --no-trust-filters available for non interactive)

"rtk trust" (--yes available for non interactive)

"rtk trust" on global + scoped config

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test — fmt clean, zero clippy warnings, 2310 unit + 8 integration tests pass (21 new tests added)
• Manual testing: rtk <command> output inspected — bare jj log / stat /usr/bin/* / gcc --version transparently rewritten to rtk … by the hook; stat/du emit [see remaining: tail -n +N …] and following it recovers the omitted lines byte-exact; RTK_TEE=0 shows full raw with no truncation marker; git status unchanged and cd / reserved names / RTK_NO_TOML=1 correctly pass through
• Human testing : Ran rtk commands and created new toml filter to test security gates, all rewrite verified in "rtk gain --history" and tested the recall with the tee_tail hint command
• Human testing : Tested all UX with interactive or non interactive guards (rtk init -g, rtk trust/untrust)

[aeppling]

Notes

This is a breaking changes for users, custom filters are now gated by rtk trust to avoid security issues.
Need a way to properly inform users as release lands

For now no breaking changes triggered for sem ver but a message in rtk gain

[pszymkowiak]

Reviewed and verified locally (built from the branch): cargo fmt clean, cargo clippy --all-targets zero warnings, toml-orphan + trust tests pass.

Part 1 (hook wiring) — functional matrix confirms it: jj / jq / ssh / mise run / just / task rewrite to rtk …; absolute paths, redirects, compound (jj diff && jq . → rtk jj diff && rtk jq .), and env-prefix are preserved; cd / reserved / unknown commands are left alone; RTK_NO_TOML=1 disables the TOML fallback while keeping the static RULES.

Part 3 (trust gating) — verified end-to-end: an untrusted .rtk/filters.toml does not activate rewrites until rtk trust, then it does.

Part 2 (tee tail hint) — verified: truncated output carries [see remaining: tail -n +N <teefile>], following it recovers the omitted lines, and RTK_TEE=0 shows full raw with no truncation marker.

Security is sound: collect_match_patterns aggregates only trusted (SHA-256) + built-in filter patterns, so a malicious project filter can't inject a rewrite; reserved / rust-handled commands are excluded from the fallback.

Non-blocking nits: the tee file drops the final trailing newline, so hint recovery is byte-exact except that last \n (looks pre-existing). The ~3–8 ms one-time cost on the Unsupported path justifies the documented match-only RegexSet follow-up.

LGTM 👍