fix: bound connecting in the shared HTTP client and log setup retries

Description of the Change

Problem

GetHTTPClient builds its own http.Transport with only TLS (and proxy) settings, so it has no dial timeout and no TLS handshake timeout. When a host never answers, each attempt waits for the OS SYN timeout (about 2m10s on Linux). Flow setup retries 3 times through retryTransient, which logs nothing per attempt, so a flow with an unreachable LLM endpoint sits silently for about 7 minutes. HTTP_CLIENT_TIMEOUT (#205) caps the whole request but doesn't help with this.

Solution

• pkg/system/utils.go: both transports get a net.Dialer with Timeout: 30s, KeepAlive: 30s and TLSHandshakeTimeout: 10s, the same values as http.DefaultTransport and the reverse proxy in pkg/server/router.go. Proxy and TLS settings are unchanged.
• No ResponseHeaderTimeout and no change to Client.Timeout: slow reasoning answers are still bounded only by HTTP_CLIENT_TIMEOUT.
• pkg/providers/retry.go: retryTransient logs one warning per failed attempt with the attempt number and the error. Deterministic errors and cancellation return as before.

Closes #460

Type of Change

• 🐛 Bug fix (non-breaking change which fixes an issue)

Areas Affected

• External Integrations (LLM/Search APIs)

Testing and Verification

Test Configuration

PentAGI Version: 2.2.0-ce (843dfeb) Docker Version: 29.9.1 Host OS: Linux LLM Provider: OpenAI (unreachable endpoint, no requests reach the API) Enabled Features: none

Test Steps

1. go test ./pkg/system/ ./pkg/providers/ -race
2. go vet, golangci-lint run ./pkg/system/... ./pkg/providers/...
3. Live: point OPEN_AI_SERVER_URL at a host that drops SYNs and create a flow, with and without the patch.

Test Results

• Unit tests (no network needed; the unreachable host is a loopback listener with a full accept queue, so Linux drops the SYNs):
  • an unreachable endpoint and an unreachable proxy fail within the dial timeout
  • a server that accepts but never answers the TLS handshake fails within the handshake timeout
  • a slow answer is not cut off, and requests still go through the configured proxy
  • the transports keep the CA pool, InsecureSkipVerify, the proxy, and Client.Timeout; there is no response header timeout
  • cancel during the dial returns at once
  • retryTransient logs once per failed attempt, and not for deterministic errors or cancellation
  • The connect and log tests fail on the current code.
• Setup call against an unreachable host:

• Live flow: before, the flow failed after about 7 minutes with only the final error. After, it failed at 1m40s with three call failed warnings first.
• With real defaults: a server that answers after 60s still succeeds, and HTTP_CLIENT_TIMEOUT=2 still cuts off a 5s answer.

Security Considerations

None. TLS verification and the custom CA pool are unchanged.

Performance Impact

None for reachable hosts. An unreachable host now fails each attempt in 30s instead of about 2 minutes.

Documentation Updates

None.

Deployment Notes

None. No new settings.

Checklist

• I have added tests to cover my changes
• All new and existing tests pass
• I have run go fmt and go vet (for Go code)
• Changes are backward compatible

Additional Notes

The bounds are fixed values, like in router.go. If you'd rather make the connect timeout configurable next to HTTP_CLIENT_TIMEOUT, I can add that.

HUMAN:

My very first PentAGI flow sat for about 7 minutes before failing, and docker logs pentagi showed nothing the whole time. It turned out my containers had no network, and the shared HTTP client has no connect timeout, so each of the 3 setup attempts waited on the OS's roughly 2-minute TCP timeout. I added a 30s connect timeout and a 10s TLS handshake timeout, the same values router.go already uses, plus a warning for each retry. I left response time unbounded on purpose, because slow reasoning calls are legitimate. If you'd like the timeouts configurable next to HTTP_CLIENT_TIMEOUT, I'm happy to add that.