Skip to content

Run SDK evolution across all vendor packages#9

Closed
ebarti wants to merge 8 commits into
mainfrom
sdk-evolution-all-packages-20260622
Closed

Run SDK evolution across all vendor packages#9
ebarti wants to merge 8 commits into
mainfrom
sdk-evolution-all-packages-20260622

Conversation

@ebarti

@ebarti ebarti commented Jun 22, 2026

Copy link
Copy Markdown
Owner

Summary

Runs the local SDK evolution workflow across the full default vendor package set instead of only google-antigravity.

This PR supersedes the narrower Antigravity-only proof by carrying forward the candidate API diff hardening and adding an all-package report/lock update.

Documentation Added

Adds docs/sdk-evolution-agent-design.md before implementing changelog and behavior-probe changes. The design document defines:

  • report-only, real-runtime, all-package, implementation, and draft-PR operating modes,
  • primary changelog and release-note sources for Claude, Codex, Codex CLI, and Antigravity,
  • why API signatures are necessary but not sufficient,
  • behavior probe tiers for unit, installed-SDK contract, and optional live checks,
  • deterministic gates for missing diffs, missing release-note evidence, failed behavior probes, reviewer rejection, and recursive self-adaptation,
  • visuals for the full workflow, runtime analysis sequence, behavior probes, and self-adaptation,
  • caveats, concerns, and alternatives considered.

docs/sdk-evolution-agent.md now links to this design doc from the user-facing run guide.

Agent-Generated Evidence

The committed report was generated through agent-runtime-kit using the real antigravity-agent-sdk runtime from a temporary pre-update checkout, with no --package filter:

env -u UV_EXCLUDE_NEWER uv run python -m examples.sdk_evolution_agent \
  --runtime antigravity-agent-sdk \
  --refresh-preview \
  --report-dir /tmp/ark-sdk-evolution-all-report

Committed report:

reports/sdk-evolution-all-packages/20260622T091555Z/report.md

The resolver found two update candidates:

  • claude-agent-sdk: 0.2.96 -> 0.2.106
  • google-antigravity: 0.1.2 -> 0.1.4

Codex package status:

  • openai-codex: already current at 0.1.0b3
  • openai-codex-cli-bin: no resolver update; the lock remains on 0.137.0a4, which is newer than PyPI's stable latest metadata field

Implementation

  • Updates uv.lock for claude-agent-sdk to 0.2.106.
  • Keeps the Antigravity lock update to google-antigravity 0.1.4.
  • Replaces the Google-only proof artifact with an all-package proof artifact.
  • Uses resolver update targets from uv lock --dry-run -P ... for candidate API snapshots, avoiding false downgrade diffs when PyPI stable latest is lower than a locked prerelease.
  • Adds regression coverage for resolver-target candidate snapshotting.

Verification

  • env -u UV_EXCLUDE_NEWER uv run ruff check .
  • env -u UV_EXCLUDE_NEWER uv run mypy
  • env -u UV_EXCLUDE_NEWER uv run pytest (99 passed, 3 skipped)
  • env -u UV_EXCLUDE_NEWER uv lock --check
  • env -u UV_EXCLUDE_NEWER uv lock --dry-run -P claude-agent-sdk -P openai-codex -P openai-codex-cli-bin -P google-antigravity (No lockfile changes detected)
  • env -u UV_EXCLUDE_NEWER AGENT_RUNTIME_KIT_LIVE=1 AGENT_RUNTIME_KIT_LIVE_PROVIDER=all uv run pytest tests/test_live_smoke.py -q -s (3 passed)

Antigravity Comparison With PR #8

Compared with PR #8 (sdk-evolution-proof-20260622):

  • src/agent_runtime_kit/adapters/antigravity.py: no diff.
  • tests/test_antigravity_adapter.py: no diff.
  • google-antigravity lock state: unchanged at 0.1.4.
  • google-antigravity API diff artifact: identical to PR Update google-antigravity from SDK evolution report #8.
  • Difference is scope: this PR replaces the Google-only report with an all-package report and adds the Claude lock update plus resolver-target candidate selection.

Safety

  • No auto-merge.
  • No unsupported credential scraping.
  • Freshness cutoffs are explicitly removed for SDK evolution commands.
  • Candidate SDK inspection uses temporary isolated virtualenvs.

Codex Model Policy

Codex-backed SDK evolution stages now explicitly run with model=gpt-5.5 and reasoning_effort=xhigh. The override is scoped to codex-agent-sdk only; Claude and Antigravity keep provider-native model selection because gpt-5.5 is not a portable model id across adapters.

@ebarti

ebarti commented Jun 22, 2026

Copy link
Copy Markdown
Owner Author

Superseded by the split PRs: #10 contains the SDK evolution example/tooling fixes; #11 contains only the generated SDK update run artifacts and lockfile changes, stacked on #10.

@ebarti ebarti closed this Jun 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant