Where: packages/safe-bash/src/commands/text-programs/awk-runtime.ts:296 — string concatenation is string(this.budget.check(this.asText(left) + this.asText(right))): a full string copy per concat, with budget.check bounding only the result size (≤ 32 MB maxBufferBytes) and no budget.step() proportional to bytes copied. The enclosing for loop charges 1 step per iteration regardless of how large the accumulated string is.
PoC (Node 22, default limits):
echo | awk '{ s=""; for(i=0;i<100000;i++) s = s "x"; print length(s) }'Measured: 100k iters → 2.3–4.9 s; 200k → > 20 s; 1M → maxCpuMs at 30 s. Classic O(n²): iteration i copies i bytes ⇒ Σ = n²/2 byte-copies. The step budget (5M) would permit 5M iterations ≈ 12.5 TB of byte copying (hours) if the CPU clock didn't intervene; the only real bound is the 30 s wall/CPU deadline, and each concat of a multi-MB string is a synchronous stretch between checkpoints.
Impact: (d) — a one-line awk program pins a core for the full wall-clock window per exec, repeatable. Same under-priced-step class as v2 M4 (awk sprintf), #634 (ERE), and #646 (join -o), but on the single most common awk operation — string accumulation — which real scripts do constantly (building output in a variable). The in-repo model for the fix already exists: the yq encoder and the arrayJoin/rebuild paths charge budget.step(bytes).
Fix: charge budget.step(Math.max(1, Math.ceil(resultBytes / 64))) (or similar) in the concat branch so byte-work is priced; optionally intern/rope the accumulator to make repeated concat O(1) amortized. Either way the loop should hit a catchable ProgramError/step limit in well under a second, not ride the wall clock.
Where:
packages/safe-bash/src/commands/text-programs/awk-runtime.ts:296— string concatenation isstring(this.budget.check(this.asText(left) + this.asText(right))): a full string copy perconcat, withbudget.checkbounding only the result size (≤ 32 MBmaxBufferBytes) and nobudget.step()proportional to bytes copied. The enclosingforloop charges 1 step per iteration regardless of how large the accumulated string is.PoC (Node 22, default limits):
Measured: 100k iters → 2.3–4.9 s; 200k → > 20 s; 1M →
maxCpuMsat 30 s. Classic O(n²): iteration i copies i bytes ⇒ Σ = n²/2 byte-copies. The step budget (5M) would permit 5M iterations ≈ 12.5 TB of byte copying (hours) if the CPU clock didn't intervene; the only real bound is the 30 s wall/CPU deadline, and eachconcatof a multi-MB string is a synchronous stretch between checkpoints.Impact: (d) — a one-line awk program pins a core for the full wall-clock window per exec, repeatable. Same under-priced-step class as v2 M4 (awk
sprintf), #634 (ERE), and #646 (join-o), but on the single most common awk operation — string accumulation — which real scripts do constantly (building output in a variable). The in-repo model for the fix already exists: the yq encoder and thearrayJoin/rebuildpaths chargebudget.step(bytes).Fix: charge
budget.step(Math.max(1, Math.ceil(resultBytes / 64)))(or similar) in theconcatbranch so byte-work is priced; optionally intern/rope the accumulator to make repeated concat O(1) amortized. Either way the loop should hit a catchableProgramError/step limit in well under a second, not ride the wall clock.