Can recorded trajectories prepare a faithful execution starting point?¶

Main conclusions¶

Seven pinned trajectory formats each pass 60/60 prefix checks. Prepare-only executes no tools and leaves the workspace unchanged. Most formats show separated timing clusters: lower-cluster medians are about 6–7 ms and higher-cluster medians about 11–16 ms. This supports binding tasks to historical tool observations; fidelity of subsequent model execution remains unmeasured.

Motivation¶

Training and reproduction need tasks to continue at an accurate historical boundary. Missing tool observations or including a future action changes the model's context; executing tools during preparation may introduce additional side effects. These properties determine whether recorded trajectories provide reliable task inputs.

Experiment design¶

Each adapter receives twenty synthetic native-format trajectories, each with two tool batches, using after-step=1 and prepare-only. Three repetitions shuffle adapter/task order with a fixed seed, giving 420 attempts. The table pins format profiles separately from installed Agent CLI versions.

Valid samples must preserve the source digest, exact command arguments and historical observations for the complete first batch, exclude the second action, report replayed_tool_calls=0 with no Agent started, and retain both the pre-existing workspace file's contents and the file inventory. An independent publication audit rechecks every retained command, native prefix and workspace. Failures are counted separately rather than assigned zero latency.

The environment is Linux/x86_64 on an AMD Ryzen 7 9700X with kernel 7.2.8-200.fc44. Processes are pinned to CPUs 0 and 1; no benchmark-specific host memory cap is imposed. Host caches are not cleared and replay has no warmups. All valid slow samples remain, with no post-hoc interference exclusion. Timing spans CLI launch through prepare-only exit, including launch and teardown but excluding input generation and post-run audit. There are no model requests or tool executions.

Native Agent resume and RL pipeline prefix preparation are relevant alternatives, but neither is a matched-trajectory, matched-budget timing control here. This experiment validates preparation for pinned formats, without measuring full session restoration, identical model next actions or rewards.

Data and analysis¶

Measured on 2026-10-06. N=60 per adapter, all passing; units are ms. Separated clusters have their own shares and medians. P95 describes the entire adapter distribution as an observation reference, with no stable tail-latency claim.

Mini-SWE-Agent does not trigger the split rule, so its row reports the full-sample P50. Splitting is descriptive: each cluster contains at least 10% of samples; the largest adjacent gap is at least 20% of the overall median and exceeds three times the median adjacent gap; cluster medians differ by at least 1.5×. This does not establish the cause of slower samples. All valid slow samples remain, so preparation should not be budgeted as a fixed few milliseconds.

Scope¶

These results support structure, argument and observation fidelity for pinned short prefixes. Real sessions, newer CLIs, long prefixes, token costs, macOS, remote-connection recovery and subsequent model execution remain unmeasured. Tool replay and actual environment restoration need separate correctness and cost checks.

Downloads and reproduction¶

Derived data CSV · Artifact and audit provenance CSV · Evidence source summary · Comparison method · Runner manual