feat(refacto-core): binary hook w/ native cmd exec + streaming

[aeppling]

From the idea of #536 to implement with current code structure and standardize with codebase (no duplication/divided flow)

Summary

Binary Hook Engine

• Native rtk hook claude|cursor|copilot|gemini handlers replacing bash scripts
• Claude Code JSON protocol (tool_name/tool_input/updatedInput)
• Copilot VS Code JSON + CLI string toolArgs auto-detection
• Cursor/Gemini dedicated protocol handlers
• rtk hook check [--agent] <command> dry-run testing
• 890 lines of bash deleted (4 hook scripts removed)
• rtk init migrates legacy script hooks to binary automatically

Permission System

• Permission verdict engine: Ask/Allow/Deny/Default (Deny > Ask > Allow > Default)
• Loads deny/ask/allow rules from Claude Code settings.json (project + global)
• Selective permissionDecision — omitted for Ask/Default so updatedInput works in bypassPermissions
• Hook audit logging via RTK_HOOK_AUDIT=1

Streaming & Execution Infrastructure

• exec_capture() standardized .output() wrapper → CaptureResult{stdout, stderr, exit_code} — migrated across 22
  modules
• ChildGuard RAII zombie process prevention
• FilterMode enum: Streaming, Buffered, CaptureOnly, Passthrough
• StdinMode enum: Inherit, Filter, Null
• Stdout protocol safety: writeln!(io::stdout()) replacing println!

rtk pipe Command

• Stdin filter dispatch with 20+ backends via resolve_filter(name)
• auto_detect_filter() from first 1KB of input
• --filter <name> and --passthrough modes

rtk proxy Streaming

• Progressive output via threaded pipe readers with immediate flush

Discovery & Registry

• Shell prefix peeling: noglob, command, builtin, exec, nocorrect
• Trailing redirect stripping: 2>&1, 2>/dev/null preserved through rewrite
• Lexer-based has_heredoc() and strip_quotes()
• New rules: gt (Graphite), python3 -m pytest, python -m pytest, pip show, uv pip install

Git

• Multi-remote branch dedup for git branch -a
• exec_capture migration preserving SSH/GPG signing

Init System

• Binary hook registration in settings.json
• Atomic file writes: NamedTempFile + persist()
• Multi-agent init: Claude, Copilot, Cursor, Gemini

Test plan

• 1583 tests pass (cargo test --all)
• Hook routing retest
• 20+ Chained commands

Should fix :

ISSUE #154 : migrate bash hook to Rust subcommand
ISSUE #222 : proxy streaming for long-running commands
ISSUE #530 : strip trailing stderr redirects before pattern matching
ISSUE #532 : env var prefix and cd chaining in rewrite
ISSUE #897 : subprocess memory leak / zombie process prevention (ChildGuard)
Closes #968 : EAGAIN posix_spawn resource exhaustion (binary hooks)
ISSUE #886 : RTK bypasses Claude Code permissions (permission verdict system)
ISSUE #712 : rtk hook subcommands added
ISSUE #918 : env-var-prefixed exclude_commands (partial: -h flag conflict not addressed)
ISSUE #893 : updatedInput in bypassPermissions mode (partial: default/ask fixed, allow edge case remains)
ISSUE #928 : python3 -m pytest/mypy rewrite (partial: heredoc/script.py not covered)
ISSUE #361 : 5 of 6 hook bugs addressed (streaming, compound commands, conservative routing)
ISSUE #295 : uv sync/pip patterns (~5% of requested scope)

BREAKING CHANGE

For claude code user, they'll need to re-install with "rtk init -g", because claude code now use native binary instead of rewrite

Old script supported until June 13th 2026

[aeppling]

Removed strip_quote from last release requirements, will be added back for this

[aeppling]

not ready for review, need fix for a full standardized flow

[aeppling]

@ahundt

[pszymkowiak]

Tested on multipass VM (Ubuntu 24.04, aarch64): 106/110 passed, avg 85% savings. All 15 ecosystems pass. 1450 unit tests pass. Same 2 pre-existing failures as other PRs (clippy warnings + diff test bug).

P0 — must fix:

1. stream.rs:190-198 — Silent 1MB truncation in Buffered/CaptureOnly modes
Lines beyond RAW_CAP (1 MiB) are silently dropped. cargo test with verbose output or pytest with many tests easily exceeds 1 MiB. The filter receives truncated input and produces wrong output (e.g. missing test summary). Add a warning when the cap triggers:

} else if !capped {
    capped = true;
    eprintln!("[rtk] warning: output exceeds 1 MiB — filter input truncated");
}

2. main.rs:2031 — Commands::Run comment says "with token tracking" but has no tracking
No TimedExecution, no timer.track(). All rtk run commands show 0 in rtk gain. Either add tracking or remove the misleading comment.

P1 — important:

3. stream.rs — Streaming, Buffered, Passthrough variants have zero production callers
22 clippy dead_code warnings from this PR. Either gate with #[cfg(test)] or wire to at least one real command.

4. hook_cmd.rs:288 — run_claude silently swallows JSON parse errors
run_copilot (line 53) logs to stderr: "[rtk hook] Failed to parse JSON input: {e}". run_claude just returns Ok(()). Match the copilot pattern.

5. runner.rs — skip_filter_on_failure stderr ordering changed
Previously stderr was printed after filtering; now it streams live during execution. Confirm this is intentional and update cmds/README.md (lines 65-68).

Fix P0s + P1.4 (one-liner) and it's ready.

[bruisedsamurai]

@pszymkowiak Merging this anytime soon?

[bruisedsamurai]

Hey @pszymkowiak I might be impatient but I will appreciate if you review,

[bruisedsamurai]

@aeppling press the button