land #8861: deliver piped stdin to -p

[proggeramlug]

Lands #8861 (fix(stdin): deliver piped stdin to -p) with a missing GC root fixed.

The PR itself fixes three stacked defects that together made echo hi | claude -p "…" print nothing while exiting 0. Its diagnosis holds up on review:

1. process.stdin.once(…) was never lowered — perry-hir matched only ("stdin","on") | ("stdin","addListener"), so once fell through to the generic member-call path and never registered with the fd-0 reader. Decisive, because the print-mode reader awaits race(stdin.once("end"), timeout(3000)); with once dropped the end half could never win, and the timer is unref'd so nothing kept the loop alive.
2. Only one stdin.on("end") listener survived — they shared readline's single-slot CLOSE_CALLBACK; the bundle registers three. Now a list fired in registration order. Verified the pump uses std::mem::take, so each fires exactly once and all fire in order.
3. Pull-mode bytes were discarded — on("readable") + read() set neither the raw nor the flowing flag, so bytes were routed to readline's line queue that read() never drains.

Fix applied while landing: a missing GC root

gc_runtime_root_holders flagged the new STDIN_END_CALLBACKS: Mutex<Vec<i64>> — it stores closure pointers and no registered scanner reached it.

This is a real missing root, not a classification gap. scan_readline_roots_mut already visits DATA_CALLBACKS, KEYPRESS_CALLBACKS, READABLE_CALLBACKS and the single-slot CLOSE_CALLBACK this list replaces — the new list was simply not added. Its closures are reachable only from there between registration and EOF, so a collection in that window would leave stale pointers that the pump then calls. Added &STDIN_END_CALLBACKS to the scanner.

That is exactly the failure mode CLAUDE.md describes for an unrooted runtime-side cache: it goes bad at collection #0 and stays bad.

Also applied

• check_file_size: readline/mod.rs reached 2049 lines. Extracted its #[cfg(test)] mod tests (262 lines) to readline/mod_tests.rs; now 1796.
• cargo fmt --all.

Note on once semantics

The lowering routes once to the same registry as on. once("end") gets true one-shot behaviour because the pump takes the list, but once("data") / once("readable") behave like on — a deviation from Node that the PR documents inline rather than hides. That is a net improvement (previously once did nothing at all), but it is a residual worth its own issue.

Validation

• all 30 lint-job gates pass
• new process_stdin_once_lowering test: 2 passed
• perry-stdlib 122, perry-hir 339, perry-runtime 2702 — all 0 failed

Summary by CodeRabbit

• Bug Fixes
  • Fixed process.stdin.once(...) so EOF and other stream events are delivered correctly.
  • Improved stdin handling for end, close, readable, and data events.
  • Preserved multiple end/close callbacks and invoked them in registration order.
  • Improved paused and flowing input modes, including cooked-mode data delivery.
  • Ensured pending stdin callbacks keep the event loop active until processed.
• Tests
  • Added comprehensive coverage for stdin events, lifecycle cleanup, pause/resume behavior, and readline input handling.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: e3232c8c-68d8-45c2-b7eb-11310025cd95

📥 Commits

Reviewing files that changed from the base of the PR and between 926d957 and d771dc7.

📒 Files selected for processing (6)

• crates/perry-hir/src/lower/expr_call/module_class_static.rs
• crates/perry-hir/tests/process_stdin_once_lowering.rs
• crates/perry-stdlib/src/readline/mod.rs
• crates/perry-stdlib/src/readline/mod_tests.rs
• crates/perry-stdlib/src/readline/pump.rs
• crates/perry-stdlib/src/readline/test_support.rs

---

📝 Walkthrough

Walkthrough

process.stdin.once(...) now uses the stdin listener lowering path. Readline now supports multiple end/close callbacks, readable pull mode, expanded cooked-input routing, ordered EOF dispatch, state cleanup, and comprehensive tests.

Changes

Stdin listener support

Layer / File(s)	Summary
Lower stdin once-listeners crates/perry-hir/src/lower/expr_call/module_class_static.rs, crates/perry-hir/tests/process_stdin_once_lowering.rs	process.stdin.once(...) now lowers to ProcessStdinOn. Regression tests cover once, on, and addListener.
Manage stdin listener state crates/perry-stdlib/src/readline/mod.rs	Readline stores multiple end and close callbacks, tracks readable pull mode, routes cooked input, and updates listener removal and GC scanning.
Dispatch and reset stdin callbacks crates/perry-stdlib/src/readline/pump.rs, crates/perry-stdlib/src/readline/test_support.rs, crates/perry-stdlib/src/readline/mod_tests.rs	The pump invokes pending EOF callbacks in registration order and keeps the event loop active while callbacks remain. Reset logic and tests cover listener, queue, lifecycle, input, and raw-mode behavior.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant ProcessStdinOn
  participant ReadlineListeners
  participant ReadlinePump
  ProcessStdinOn->>ReadlineListeners: register stdin listener
  ReadlineListeners->>ReadlinePump: enable pull mode or EOF observation
  ReadlinePump->>ReadlineListeners: deliver input and end/close callbacks

Loading

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch land-8861-stdin

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}