[chaliy]

Requested by Михайло · project thread

What changed

• Embedded python3 (Monty) can be called as often as a script likes. Each call now pays the shared work budget for the heap it actually uses (one unit per 64 bytes past the run's high-water mark) instead of reserving max_memory / 64 units up front. Repeated large runs still share one budget (TM-DOS-096 holds: the regression test now uses two 12 MiB runs against a ~19 MiB budget).
• The last expression is no longer echoed (breaking: python3 -c "2 + 2" prints nothing, like CPython for -c, files and stdin). Every program ends with a None statement so Monty's final value is always None.

Why

Found reviewing a downstream integration (harnx), reproduced on 0.20.3:

• With defaults, the reservation was 1M of 100M units, so the 96th python3 -c pass in one script failed with work units exhausted.
• Monty converts the module's final value to a host object outside its memory and time limits. x = [x, x] thirty times as the last line expands to 2^30 leaves (~500 MiB) and ignores the timeout. Bashkit can't bound that conversion, so it must never happen.

Before / After

Before:

$ bashkit -c 'for i in $(seq 1 100); do python3 -c pass || { echo "failed at $i"; break; }; done'
Error: resource limit exceeded: execution budget exhausted: work units (100663793 used, max 100000000)
$ timeout 20 bashkit -c "python3 -c 'x=[1]
for i in range(30): x=[x,x]
x'"; echo $?
124

After: 300 calls in one script succeed (many_small_invocations_fit_in_one_script); the shared-graph program exits 0 immediately (test_large_shared_final_value_finishes); a heavy-memory loop still exhausts a small budget (memory_used_is_charged_to_the_work_budget, verified failing with the charge removed).

Risk

• Medium
• Users relying on REPL-style echo get no output; docs (python.md, knowledge) now say to use print(). Code ending in a trailing backslash is left unmodified so it stays a SyntaxError.

Checklist

• Tests added or updated
• Backward compatibility considered

---

Generated by Claude Code

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	bashkit	739ac73	Commit Preview URL Branch Preview URL	Nov 05 2026, 03:55 AM

[bashkit-pr-reviewer]

Review verdict: approve-worthy, no findings

This PR fixes two real bugs in the Python builtin and the fix approach is sound. Charging actual Monty heap high-water (charge_memory_peak, 64 bytes/work unit) instead of reserving max_memory / 64 up front removes the ~95-call cap on trivial python3 -c 'pass' invocations while keeping TM-DOS-096 enforced — the updated tests prove both directions (300× pass succeeds; repeated 12–16 MiB heaps still exhaust the shared work budget). Appending \nNone\n so Monty's final value is always None closes the ~500 MiB shared-list host-expansion hole, and matches CPython (-c/file/stdin never echo the last expression); the trailing-backslash guard preserves the syntax-error path and the Complete(_) arm already ignores the result, so nothing user-visible is lost.

Findings: none — no P0/P1/P2. Ordering in charge_memory_peak (record high-water before consume_work, so a failed charge isn't double-counted) and the monotonic charged_memory (frees don't refund, regrowth to the same peak costs nothing) are both correct for a peak-based charge.

Context read beyond the diff: AGENTS.md, knowledge/runtimes/python-builtin.md, knowledge/log.md, crates/bashkit/docs/python.md, full run_python loop (budget entry, OsCall/FunctionCall/Complete arms), BudgetTracker::{consume_work, charge_memory_peak, check_vm, on_grow}, unit tests (test_inline_expression_is_not_echoed, test_large_shared_final_value_finishes, test_trailing_backslash_stays_a_syntax_error), integration tests (execution_budget_tests.rs, python_integration_tests.rs, python_security_tests.rs incl. class-escape module), and tests/spec_cases/python/python.test.sh. Verified the PR's claims against project rules (regression tests added, knowledge + docs updated in the same change, decisions recorded as comments).

Update on latest push (head 739ac73, rebased onto 2f35ec2): re-checked out and re-verified — same 8 files, same +143/−42 shape, same logic (high-water charge, None suffix + backslash guard, same regression tests). No new changes to review; earlier conclusions stand.

Not reviewed: nothing material. Diff reported truncated: false. As usual I could not build or run tests (no cargo in this shell) — I reasoned from the code, and LimitedTracker::current_memory lives in the external Monty crate so I took its byte-count semantics from the name and usage.