fix: isolate claude-code sessions via stream-json session_id

[codefromthecrypt]

Summary

claude-code shares a persistent subprocess that previously conflated all sessions into one. This caused a lot of problems, notably tests would include state from other tests. This could add pressure to restart the subprocess, rather than solve the bookkeeping problem.

This implements session_id to the stream-json protocol and removes the messages_sent bookkeeping as sessions are handled internally to claude.

This is standard practice for stream-json SDKs: Each sends only the new user message per turn with a session_id field, and the CLI maintains conversation context internally:

• anthropics/claude-agent-sdk-python — session_id defaults to "default"
• guess/claude_code (Elixir) — session_id defaults to "default"
• severity1/claude-agent-sdk-go — session_id on every StreamMessage

This also updates providers.rs integration tests to use a unique session ID per test, and support running with claude-code and codex CLI providers.

Type of Change

• Bug fix
• Refactor / Code quality
• Tests

AI Assistance

• This PR was created or reviewed with AI assistance

Testing

$ OLLAMA_HOST=http://localhost:11434 cargo test --test providers -- test_claude_code_provider test_codex_provider test_ollama_provider --nocapture
   Compiling goose v1.23.0 (/Users/codefromthecrypt/oss/goose-2/crates/goose)
    Finished `test` profile [unoptimized + debuginfo] target(s) in 2.08s
     Running tests/providers.rs (target/debug/deps/providers-edd2a25b142813c2)

running 3 tests
=== codex::model_listing ===
[crates/goose/tests/providers.rs:288:9] &models = [
    "gpt-5.2-codex",
    "gpt-5.2",
    "gpt-5.1-codex-max",
    "gpt-5.1-codex-mini",
]
===================
=== claude-code::model_listing ===
[crates/goose/tests/providers.rs:288:9] &models = [
    "sonnet",
    "opus",
]
===================
=== Ollama::model_listing ===
[crates/goose/tests/providers.rs:288:9] &models = [
    "all-minilm:33m",
    "all-minilm:l6-v2",
    "devstral:latest",
    "gemma3:latest",
    "llama3.2:latest",
    "nomic-embed-text:latest",
    "qwen2.5-coder:32b",
    "qwen2.5:0.5b",
    "qwen2.5:latest",
    "qwen3-vl:2b",
    "qwen3-vl:latest",
    "qwen3:0.6b",
    "qwen3:1.7b",
    "qwen3:4b",
    "qwen3:latest",
]
===================
=== claude-code::basic_response === Hello! 👋 How can I help you today?
=== claude-code::context_length_exceeded_error ===
[crates/goose/tests/providers.rs:241:9] &result = Err(
    ContextLengthExceeded(
        "Prompt is too long",
    ),
)
===================
test test_claude_code_provider ... ok
=== codex::basic_response === Hello!
=== Ollama::basic_response === Hello! 😊 How can I assist you today?
=== codex::context_length_exceeded_error ===
[crates/goose/tests/providers.rs:241:9] &result = Err(
    ContextLengthExceeded(
        "Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.",
    ),
)
===================
test test_codex_provider ... ok
=== Ollama::tool_usage === test-uuid-12345-67890
=== Ollama::image_content === the image displayed contains a simple text message on a white background. it reads:  
**"hello goose! this is a test image."**  

the text is presented in a clean, sans-serif font with black lettering. the design is minimalistic, with no additional graphics, logos, or decorative elements. the message appears to be a straightforward test payload for validation purposes.
=== Ollama::context_length_exceeded_error ===
[crates/goose/tests/providers.rs:241:9] &result = Ok(
    (
        Message {
            id: None,
            role: Assistant,
            created: 1770690098,
            content: [
                Text(
                    Annotated {
                        raw: RawTextContent {
                            text: "no",
                            meta: None,
                        },
                        annotations: None,
                    },
                ),
            ],
            metadata: MessageMetadata {
                user_visible: true,
                agent_visible: true,
            },
        },
        ProviderUsage {
            model: "qwen3-vl",
            usage: Usage {
                input_tokens: Some(
                    79,
                ),
                output_tokens: Some(
                    189,
                ),
                total_tokens: Some(
                    268,
                ),
            },
        },
    ),
)
===================
test test_ollama_provider ... ok

test result: ok. 3 passed; 0 failed; 0 ignored; 0 measured; 13 filtered out; finished in 56.61s


============== Providers ==============
✅ Ollama
✅ claude-code
✅ codex
=======================================

Related Issues

• feat: MCP support for agentic CLI providers #6972 follows up with MCP support for CLI providers

[codefromthecrypt]

this only applies to CLI validation, not the ndjson itself.

[Copilot]

Pull request overview

This PR updates the claude-code CLI provider to use a session_id in the stream-json protocol so a single persistent subprocess can correctly isolate multiple concurrent/serial Goose sessions (and avoids leaking state between tests).

Changes:

• Add session_id to the claude-code stream-json input payload and remove the previous messages_sent bookkeeping.
• Update provider integration tests to support CLI providers (claude-code, codex) and use per-test session IDs where needed.
• Refactor CLI provider from_env construction for claude-code and codex into the ProviderDef implementation.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
crates/goose/src/providers/claude_code.rs	Add session_id to stream-json input; remove messages_sent; adjust session-id handling in complete_with_model.
crates/goose/src/providers/codex.rs	Move env-based provider construction into ProviderDef::from_env.
crates/goose/tests/providers.rs	Expand integration tests for CLI providers; introduce CLI/non-CLI session handling and skip logic.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.