fix(stream): omit prior reasoning from request history

[patlux]

Summary

• stop replaying completed assistant reasoning traces to Command Code on follow-up requests
• preserve visible assistant text, completed tool calls, and matching tool results
• omit history turns that contain only private reasoning
• add unit coverage and a real two-turn Pi RPC test against the live Command Code backend

Why

The current Command Code CLI does not replay private reasoning traces as conversation history. Keeping old reasoning in later /alpha/generate requests can interfere with fresh reasoning on follow-up turns.

Verification

• npm test
• npm run test:e2e:live
  • both sequential turns emitted fresh thinking_delta events against the real backend
• npm run typecheck
• npm run format:check
• npm pack --dry-run
• git diff --check

Notes

• Uses existing Pi authentication without reading or printing credential files.
• OMP process test skips because omp is not installed on the test host.

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[github-actions]

Memory benchmark

Compared base 7d02d34 with PR head ea55ead on the same GitHub-hosted darwin runner. Lower values are better.

Metric	Base	PR	PR − Base	Change
Stable RSS	168.0 MiB ± 0.9 MiB	168.5 MiB ± 0.8 MiB	+0.5 MiB ± 1.0 MiB	+0.3%
Physical footprint	110.8 MiB ± 0.6 MiB	111.7 MiB ± 0.5 MiB	+0.8 MiB ± 0.8 MiB	+0.7%
Physical peak	127.7 MiB ± 0.7 MiB	128.0 MiB ± 0.4 MiB	+0.6 MiB ± 1.6 MiB	+0.5%

Interpretation: Differences larger than their measured MAD: Physical footprint +0.8 MiB ± 0.8 MiB.

Extension overhead above pi baseline

Metric	Base overhead	PR overhead	Difference
Stable RSS	+5.8 MiB ± 0.5 MiB	+7.0 MiB ± 1.2 MiB	+0.5 MiB ± 1.0 MiB
Physical footprint	+4.0 MiB ± 0.4 MiB	+4.7 MiB ± 0.4 MiB	+0.8 MiB ± 0.8 MiB
Physical peak	+3.3 MiB ± 0.7 MiB	+4.4 MiB ± 0.9 MiB	+0.6 MiB ± 1.6 MiB

Values are medians of 6 alternating, paired runs. The value after ± is the median absolute deviation (MAD). Each process was sampled 12 times after a 2500 ms warm-up.

Environment: pi 0.82.1, Bun 1.3.11, darwin arm64. RSS comes from ps; physical footprint and peak come from macOS footprint.

This is a comparative signal, not a pass/fail threshold. GitHub-hosted runner noise can affect absolute values.