File-based interactive Rust

(authored by agents unless marked 🧑)

• 🧑 goal: “much better interactive use by editing and loading a file or files instead of doing it REPL style and optimized for agents”
  • source: human request
• result: existing state can be inspected and mutated with Rust source files
  • reproduction: bash experiments/interactive/smoke.sh from repository root
  • measured 2026-10-06 on x86_64 Linux, rustc 1.97.1, warm dependencies
    • host rebuild: Cargo reported 0.68 s
    • inspect, mutate, inspect compilations: 136, 93, 136 ms
    • elapsed snippet execution rounded to 0 ms; later inspection took 2 ms
    • timings are individual observations under shared machine load
  • observed output: 3 orders, revenue 6249, then now 4 orders, revenue 6250
  • smoke script verifies persistent mutation, wrong context refusal, compiler diagnostics, panic recovery, client timeout, and resume
• implementation: prototype
  • host deliberately pauses at unsafe { pry::session(&mut state, dir) }
  • client atomically submits a source file; host compiles a fresh shared library and calls its exported function
    • each evaluation reserves a new library directory, even when the request filename repeats
    • this avoids the loader returning old code or overwriting a mapped library
  • host publishes output before status, using atomic renames
  • snippets use pry::snippet!(|st: &mut app::State| { ... })
  • bash experiments/interactive/bin/pry-eval SESSION FILE.rs waits up to 30 s
    • PRY_TIMEOUT_SECONDS changes client wait; timeout exits 124
    • timeout does not cancel compilation or execution; request may still execute
    • --continue resumes the host immediately
  • dependencies are reused from target/debug/deps
    • build host with RUSTFLAGS='-C prefer-dynamic'
    • smoke script supplies shared-library search paths
    • mixed old build flags caused observed colliding StableCrateId values
    • clear this experiment’s target directory after changing compiler or build flags
• limits: trusted development experiment, with an unsafe Rust ABI boundary
  • Rust Reference: “The Rust ABI offers no stability guarantees”
    • external blocks, ABI
  • matching type_name only catches obvious mismatches
    • it proves neither memory layout nor dependency/compiler compatibility
    • exact compiler, dependency artifacts, type layout and compatible Rust ABI are caller obligations
  • snippets must return without retaining the borrowed state
  • catches unwinding panics; state changes before panic remain
    • abort, process exit, memory corruption and an infinite loop can kill or block the host
  • loaded libraries remain resident, so repeated evaluation consumes memory and disk
  • stdout/stderr capture redirects process-wide descriptors
    • concurrent output may be mixed; multiple sessions are unsupported
    • descriptor-operation errors are currently unchecked
  • host filesystem errors can panic; compiler subprocess has no host-side deadline
  • dependency discovery chooses newest matching artifacts heuristically
  • Linux/Unix-specific loader and descriptor calls; no sandbox or arbitrary process attachment
• existing alternative: Evcxr for persistent standalone evaluation
  • Evcxr describes itself as “An implementation of eval() for Rust”
    • Evcxr library README
  • its REPL README calls it “A Rust REPL (Read-Eval-Print loop)”
    • REPL README
  • recommendation: evaluate Evcxr for standalone file evaluation; retain explicit host pause points for live borrowed state
    • Evcxr was researched, not benchmarked in this run
• research next: attach through explicit application control points and a versioned state boundary
  • recommendation: move untrusted evaluation to a separate process with serialized state
  • tradeoff: serialized state loses arbitrary direct access to live Rust references

Last edited: 2026-10-08