[pierreraby]

Summary

Carries #165's taste-learning opt-in (29a31de) by @patlux / Patrick Wozniak, credited as co-author. Independent main-based PR.

• Generate sends x-taste-learning: false unless COMMANDCODE_TASTE_LEARNING=1; both ZDR aliases force false.
• Address the original review concern: filter caller taste-learning headers case-insensitively, so caller headers cannot override this policy. Preserve unrelated custom headers.
• Document metadata and distinguish dedicated empty local-file fields from conversation/tool content still sent for inference.

Validation

• Original tests against main: 2 failures (default header and opt-in/ZDR policy).
• New caller-header regression fails against the original patlux implementation before hardening.
• Full npm test: PASS, stream suite 45/45, typecheck and real-host mock Pi/OMP included.
• Format and diff checks: PASS.
• Active all-server LSP: zero reported diagnostics, 1 clean / 2 inconclusive. Separate cached inferred-project resolution cascade on unchanged index.ts remains a baseline coverage gap, not claimed fixed. No unrelated index/config changes included.

No real secrets, billed requests, fresh independent review, merge or release. #165 remains open. Happy to close if @patlux prefers landing directly.

[github-actions]

Dependency Review

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

Scanned Files

None

[github-actions]

Memory benchmark

Compared base 80e7e93 with PR head 5724239 on the same GitHub-hosted darwin runner. Lower values are better.

Metric	Base	PR	PR − Base	Change
Stable RSS	107.9 MiB ± 0.3 MiB	107.3 MiB ± 0.6 MiB	-0.1 MiB ± 0.5 MiB	-0.1%
Physical footprint	61.4 MiB ± 0.7 MiB	60.5 MiB ± 0.5 MiB	-1.0 MiB ± 0.2 MiB	-1.7%
Physical peak	72.9 MiB ± 0.7 MiB	71.9 MiB ± 1.4 MiB	-1.2 MiB ± 0.8 MiB	-1.6%

Interpretation: Differences larger than their measured MAD: Physical footprint -1.0 MiB ± 0.2 MiB, Physical peak -1.2 MiB ± 0.8 MiB.

Extension overhead above pi baseline

Metric	Base overhead	PR overhead	Difference
Stable RSS	+6.6 MiB ± 0.9 MiB	+6.6 MiB ± 1.2 MiB	-0.1 MiB ± 0.5 MiB
Physical footprint	+3.5 MiB ± 1.2 MiB	+2.5 MiB ± 0.3 MiB	-1.0 MiB ± 0.2 MiB
Physical peak	+5.7 MiB ± 1.8 MiB	+4.6 MiB ± 2.5 MiB	-1.2 MiB ± 0.8 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 0.84.4, Bun 1.4.4, 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.

[pierreraby]

Follow-up on static-check coverage: a later fully scoped probe reproduced the inferred-project baseline warnings on unchanged index.ts. I also ran an expanded temporary typecheck including index, using the already-installed Pi 1.5.0 declarations (no dependency/config installation). It reports 17 index diagnostics at exactly the same positions and TS codes on clean main 26c4490 and this carry (compat/model shapes, OAuth structural types, readonly tiers, stream/option contracts). These are pre-existing static-compatibility/coverage issues, not all just missing-module noise, and this PR does not fix them. The normal project typecheck excludes index; its full tests and real Pi/OMP mock integrations remain green. A separate authoritative index-typecheck follow-up is needed; no unrelated migration was added to this focused carry.