perf(task): avoid quadratic response history diffs

[morluto]

What and why

Closes #74.

Every completed task step diffs its context against a snapshot. The top-level list diff scans every stored response against the entire prior response list, so an unchanged history of 100 responses costs 5,050 equality comparisons; appending one response costs 5,150. This makes context bookkeeping quadratic as a run accumulates responses.

Check whether the old list is an unchanged prefix, then scan only the appended tail. Tail entries still undergo the existing membership check so duplicate appends remain no-ops. Changed or reordered prefixes retain the original behavior, including the handling of mutated response records. Deep-copy snapshots and nested metadata diffs keep their existing semantics.

Synthetic Python 3.12.12 / arm64 benchmark, measured without profiler instrumentation as the median of five batches. Each cycle includes a full context snapshot and diff, with distinct stored responses containing 512-character strings and a small structured output:

Stored responses	Before	After
100	2.9 ms	0.64 ms
500	59 ms	3.0 ms
1,000	273 ms	6.3 ms

These timings measure context bookkeeping rather than complete agent runs. An approximately 246 KiB nested-metadata workload remains about 5.9 ms per cycle. The issue includes a reproduction script.

How it was tested

• Added a comparison-count regression for unchanged response histories and histories with one appended response. It fails on the original implementation and passes with the optimization, without relying on wall-clock thresholds.
• Ran the actual context test file with isolated production-module loading, package initialization and repository conftest fixtures bypassed, and five environment-dependent tests excluded: 34 passed. This isolation bypasses unavailable cloud SDK imports and does not replace the normal CI suite.
• Compared journal output with the original implementation across 58,572 cases, including appends, duplicates, deletions, reordering, and mutated response contents: identical outputs.
• git diff --check and Python syntax checks passed.

Confidence Score: 5/5

The PR appears safe to merge based on the reviewed changes.

What we checked:

• Changed responses still reach the full scan: A changed response makes the prefix comparison fail. The code then checks the full list, as it did before.

Summary

Context updates now scan only appended items when a saved list stays an unchanged prefix, avoiding repeated comparisons as histories grow. A comparison-count test covers unchanged and appended response histories.

• Context updates scan only new items when a list keeps its old prefix.

_{Reviews (1) · Last reviewed commit: "perf(task): avoid quadratic response his..."}

[earakely-scale]

Nice PR, approved. Will help get merged assuming integ tests pass

[earakely-scale]

/codebuild_run(310b0ee)

[earakely-scale]

Merging, thanks!