fix(agent): avoid closing shared Anthropic client on interrupt

[raymondyan-zhijie]

Anthropic Messages uses a shared SDK client, unlike the request-local OpenAI-wire clients protected by the existing #29507 owner-thread close discipline. The stale/interrupt poll loop was still closing and rebuilding that shared client from a non-owner thread, which can release TLS file descriptors while a worker thread is still unwinding the SSL BIO. If SQLite reuses that descriptor, pending TLS bytes can clobber a database header.

Skip shared Anthropic client close/rebuild from non-streaming stale, non-streaming interrupt, and streaming interrupt paths. The in-flight SDK call is left to its timeout or stream context cleanup while the existing OpenAI request-local close path remains unchanged.

Add regression coverage for the Anthropic-compatible paths so future interrupt/stale changes do not reintroduce shared client close/rebuild.

What does this PR do?

Related Issue

Fixes #

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

How to Test

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform:

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

For New Skills

• This skill is broadly useful to most users (if bundled) — see Contributing Guide
• SKILL.md follows the standard format (frontmatter, trigger conditions, steps, pitfalls)
• No external dependencies that aren't already available (prefer stdlib, curl, existing Hermes tools)
• I've tested the skill end-to-end: hermes --toolsets skills -q "Use the X skill to do Y"

Screenshots / Logs

[teknium1]

Thanks for tracing the shared-client ownership issue. The premise is still present on current main: non-streaming stale/interrupt paths close agent._anthropic_client at agent/chat_completion_helpers.py:717-719 and :759-761, and streaming interrupt does so at :3109-3111.

Problems

• Current main has three additional shared-Anthropic cleanup sites that this PR does not cover: mid-tool retry at agent/chat_completion_helpers.py:2834-2837, ordinary retry at :2894-2897, and stale-stream cleanup at :3080-3083. All call close() and _rebuild_anthropic_client() on the shared client.
• The non-streaming dispatch was extracted after this PR's base (agent/chat_completion_helpers.py:449-457, commit 5c5dd6b7e), so this needs a current-structure transplant rather than a direct application.

Suggested changes

• Carry the request-local Anthropic ownership discipline through all six current cleanup/interrupt paths and add regression cases for the three streaming retry/stale paths.

Automated hermes-sweeper review.

[teknium1]

When salvaging this request-local client helper onto current main, route every current shared-Anthropic cleanup site through it. Current main still closes/rebuilds the shared client in streaming mid-tool retry, retry, and stale cleanup (agent/chat_completion_helpers.py:2836, 2896, 3082) in addition to the paths changed here.

[OutThisLife]

Carried this forward in #67238 for #67142 (the direct-Anthropic cron/executions.db corruption report). Your per-request-client approach was exactly right — I reimplemented it onto current main, which had drifted since this PR:

• non-streaming dispatch was extracted into _dispatch_nonstreaming_api_request(make_client=...), so the per-request Anthropic client is threaded through a kind-aware make_client there instead of the old inline if/elif;
• streaming grew _cancel_current_stream_attempt + worker retry-cleanup sites (Anthropic streaming: stale/retry paths call _replace_primary_openai_client, causing 15-min hang on stuck streams #28161 lineage) that this branch predated.

I also pushed the invariant a bit further — the shared _anthropic_client is now never closed from inside a request at all (including the worker retry-cleanup sites), since each attempt builds a fresh request-local client; that removes the last cross-request shared-client close. #28161 no-hang is preserved via the poll-thread socket abort.

Your commits are preserved via Co-authored-by (the originals were authored as "Claude Code" — credited to you as the human contributor). Will close as superseded once #67238 lands. Thanks @raymondyan-zhijie.

[OutThisLife]

Landed via #67238 (merged) — your commits preserved via Co-authored-by. Closing as superseded. Thanks @raymondyan-zhijie.