Add a Claude Code compatibility matrix and release gate #78

Part of #73.

Blocked by: #75, #76, #77 (native dependencies). This is the release gate: it proves the combination, it does not repair it.

Objective

Prove the combined Claude Code client integration with a versioned compatibility matrix, extend the client E2E matrix with Anthropic Messages cases, and gate the "stable" claim on recorded evidence.

Authority

ADR-0014 "Evidence and release gates": every upstream path must show real-CLI multi-turn text, incremental streaming, tool call/result round trips, cancellation, and error handling; the native Anthropic path additionally proves semantic-field and SSE preservation. Publish verified CLI versions and limitations in English and Chinese, and do not claim Anthropic endorses non-Claude model compatibility.

Scope

1. Real-client matrix

• Extend scripts/e2e_gateway_client_matrix.py with Anthropic Messages request, JSON response, and streaming SSE cases.
• Run against all three upstream protocols: native Anthropic Messages, OpenAI Responses, and Chat Completions — using provider identities that are explicitly approved for live evidence.
• Cover main-session text, multi-turn history, tool call + result, parallel tools, images, compaction/resume, subagents, thinking/effort, prompt caching, cancellation, and error recovery to the extent Spike Claude Code Messages compatibility through CodexHub #74 classified them; unsupported cases are recorded as unsupported, not skipped silently.
• Record the minimum supported and the latest verified Claude Code versions, with the exact executable and platform for every claim.

2. Credential contract — explicit prerequisite

docs/agents/real-client-e2e.md requires a new Issue and a dedicated credential contract before a client is added to the live matrix; the same rule that governs xAI Grok applies here. Before any live Claude Code row is added, record on this issue:

• the credential contract shape (input file, schema id, provider identity) and the Issue that owns it;
• which provider identities are approved for live evidence and who approved them;
• that no credential material enters this issue, fixtures, logs, or commits.

If approval is not granted, leave this issue blocked and state the concrete blocker. Do not infer consent from a configured provider, and do not substitute mock-only runs for live acceptance.

3. Support table

• Produce a support table distinguishing preserved, adapted (named policy), unsupported, unknown, and provider-specific behavior, traceable to Spike Claude Code Messages compatibility through CodexHub #74's classification and Add inbound Anthropic Messages API for Claude Code clients #75's implementation.
• Unsupported features are visible in the matrix and in the client UI (card compatibility state / request diagnostics). No feature is silently claimed.
• Version drift produces an actionable report with sanitized fixtures; scheduled/report-only upstream compatibility checks become blocking only after stability criteria are approved.

4. Release gate

• Release notes and documentation state the verified Claude Code version range, the experimental scope, and the known limitations in English and Chinese.
• A maintainer records the criteria for moving from experimental to stable, and the change that would satisfy them.
• Relevant Python, Rust, and frontend suites pass at the candidate commit; the quality report runs once and remains non-blocking.

Out of scope

Protocol implementation (#75), client configuration and UI (#76), model contract (#77), and any ACP/AgentProvider work (#85).

Acceptance criteria

• Automated E2E covers the supported text/tool/SSE contract against every approved upstream protocol.
• A real pinned Claude Code CLI passes multi-turn text, incremental streaming, tool call/result round trips, cancellation, and error handling on each approved upstream path.
• The native Anthropic path separately proves semantic-field and SSE preservation.
• The support table is complete, classified, and traceable to evidence for every capability in scope.
• Unsupported features appear in the matrix and the client UI; no silent claim remains.
• Minimum and latest verified Claude Code versions are recorded with exact executables.
• Version drift generates an actionable report with sanitized fixtures.
• The credential contract and its owning Issue are recorded before the first live Claude Code matrix row; no credentials are committed or pasted into issues.
• English and Chinese release documentation state the version range, experimental scope, and limitations.
• A maintainer records the experimental → stable criteria.
• Relevant Python, Rust, and frontend suites pass on the reviewed SHA; report-only quality findings are recorded without becoming a release gate.

Required verification

Risk class: strict — release gate, live runtime evidence, security boundary. Use the repository launcher; never a bare python/pytest.

.\scripts\codexhub-python.cmd -m pytest -q --ignore=tests/test_real_client_e2e.py
.\scripts\codexhub-python.cmd -m pytest -q tests/test_real_client_e2e.py
.\scripts\codexhub-python.cmd scripts/report_quality_gates.py
Push-Location src-tauri
cargo test --locked
cargo clippy --locked --all-targets -- -D warnings
Pop-Location

Push-Location frontend
npm run build
npm run test:ui-contract
Pop-Location

Linux: ./scripts/verify-linux.sh. Real-client acceptance is a local release-operator procedure per docs/agents/real-client-e2e.md; a pass on one operating system never replaces the other.

References

• scripts/e2e_gateway_client_matrix.py, tests/test_e2e_gateway_client_matrix.py
• docs/agents/real-client-e2e.md, docs/agents/verification-policy.md, docs/agents/release.md
• docs/reviews/v0.3-agent-provider-roadmap.md
• E2E third-party: official DeepSeek Flash #497 (OpenCode Go credential-contract precedent)

Document status

ADR-0014 is committed in draft PR #556 at ed76f0a6e22d3188879504638b2943f5bb855d41, not yet merged to main. The pinned link is readable now; the document is accepted direction, not live protocol evidence.