claude-sonnet-5.5

Claude Sonnet 5.5

Generate responses through Frevana's OpenRouter Responses API (POST /openrouter/v1/responses) using anthropic/claude-sonnet-5.5.

Return the validated API response JSON unchanged for a single response. When automatic continuation occurs, return the stitched result with original responses under continuation.responses, or use --text-only for the joined text.

Model Characteristics

• Model ID: anthropic/claude-sonnet-5.5 (Anthropic Claude Sonnet 5.5 via OpenRouter).
• Best suited for well-scoped feature development, bug fixes, document work, and everyday agentic workflows.
• Use --reasoning-effort to trade off depth, latency, and cost; this model always uses thinking.

What This Skill Needs

• input or prompt via --input or --input-file (or raw JSON payload via --raw-payload-file)
• optional developer system instructions via --instructions or --instructions-file
• optional reasoning effort: --reasoning-effort <low|medium|high|xhigh|max>
• optional sampling: --temperature (provider-dependent), --max-output-tokens
• optional multi-turn conversation continuation: --previous-response-id
• optional session persistence: --session-file or --session
• optional tools: --tools or --tools-file, --tool-choice
• optional service tier: --service-tier default
• authentication: FREVANA_TOKEN (or --token) / FREVANA_API_KEY (or --api-key)
• optional attribution: FREVANA_AGENT_APP_INSTANCE_ID (or X_FREVANA_AGENT_APP_INSTANCE_ID, or --agent-app-instance-id) sent as x-frevana-agent-app-instance-id
• curl, bash, and python3

Allowed Options

• --model: anthropic/claude-sonnet-5.5 (default; alias claude-sonnet-5.5 accepted)
• --input, --prompt: user prompt string or JSON message array
• --input-file: file path containing input prompt or JSON message array
• --instructions: system developer instructions
• --instructions-file: file path containing system instructions
• --reasoning-effort: low, medium, high, xhigh, max
• --temperature: 0.0 - 2.0 (currently available only on some OpenRouter endpoints)
• --max-continuations: maximum extra requests after output truncation, 0..32, default 8 (0 disables automatic continuation)
• --max-output-tokens: integer from 1 to 128000
• --session: session name (e.g. default, chat1) saved under ~/.frevana/sessions
• --session-file: path to session file to automatically save and restore conversation context across calls
• --new-session, --new: start fresh and replace the saved session only after a successful response
• --no-session: skip loading and saving a session, even when --session-file is supplied
• --previous-response-id: ID of a previous response for multi-turn conversations
• --chat: start an interactive chat session in the terminal
• --service-tier: default
• --tools: JSON array string defining tools; availability varies by OpenRouter endpoint
• --tools-file: path to JSON file defining tools
• --tool-choice: auto or none (string or JSON object); forced tool use is unsupported
• --raw-payload-file: path to raw JSON request body
• --api-key: Frevana API key (sent as X-API-Key)
• --token: Frevana Bearer token (sent as Authorization: Bearer)
• --agent-app-instance-id: Agent App instance/team ID (sent as x-frevana-agent-app-instance-id)
• --text-only, -t: output only the response text
• --output: save full response JSON to path

The skill accepts only default service tier. The current Frevana Server checkout rejects fast and priority tiers for Claude 5.5. OpenRouter endpoints for these models require reasoning and do not support top_p or forced tool choice; temperature support varies by provider. For raw payloads, use reasoning.effort rather than a fixed reasoning-token budget. Session files store the last response ID and model, not the conversation transcript; continuation uses OpenRouter's previous_response_id. Reusing a file for a different model or loading a malformed file fails with an error; use --new-session to intentionally start over.

Automatic Output Continuation

When status=incomplete and incomplete_details.reason=max_output_tokens, the script carries the original input and generated assistant text into another request and asks for only the remaining content. For stateless OpenRouter calls, automatic continuation sends message history. If the initial request uses previous_response_id through a stateful compatible API, retain that original ID as the earlier-conversation anchor on every continuation; do not switch to the latest truncated response ID, because the current turn is already included in message history. Preserve the original store setting as well. Request instructions, model, token limit, tools, sampling, authentication, and attribution are preserved; this works with --no-session, raw non-streaming payloads, and chat mode.

Completed hosted tool items (such as web search) are replayed as context and preserved in the combined output. A completed client function call still requires a corresponding function_call_output; the script never executes client tools automatically.

For raw JSON structured-output requests (text.format or response_format), the initial request retains the format constraint. Subsequent requests generate plain-text suffixes, with the original format included in the continuation prompt. Validate the final concatenated JSON syntax (and object shape for json_object); malformed output returns nonzero with partial results preserved. Full JSON Schema validation is the caller's responsibility after continuation.

Text chunks are concatenated without inserting separators, preserving split words, JSON, and code. The stitched JSON uses the last response ID/status, combined output_text and message output, summed numeric usage fields, and continuation.responses containing every original response. Single responses remain unchanged. Continuation adds latency and billable requests; the model may still repeat content despite the continuation instruction.

Stop after --max-continuations extra requests, an empty truncated text, an unfinished hosted tool or a client tool still awaiting its result, a different incomplete reason, or an API/transport error. Preserve available partial text/JSON on stdout and in --output, return nonzero, and do not advance the saved session after a failed chain. Raw stream=true is unsupported. On success, persist the final response ID using the existing session behavior. Existing explicit response-ID sessions are separate from this stateless continuation; OpenRouter currently rejects non-null previous_response_id unless the configured API supplies its own state support.

Commands

# Complex coding and software engineering
bash <skill-path>/scripts/create_response.sh \
  --input "Implement a lock-free ring buffer queue in C++20 with atomic operations" \
  --instructions "Include comprehensive thread-safety explanations and benchmarks." \
  --reasoning-effort high \
  --output ./out/sonnet-code.json

# Tool-calling workflow
bash <skill-path>/scripts/create_response.sh \
  --input "Check deploy status for cluster 'prod-eu-west'" \
  --tools '[{"type":"function","name":"get_cluster_status","parameters":{"type":"object","properties":{"cluster_name":{"type":"string"}},"required":["cluster_name"]}}]' \
  --tool-choice auto \
  --text-only

# Multi-turn conversation with automatic session persistence (Recommended)
bash <skill-path>/scripts/create_response.sh \
  --session-file ./out/sonnet_session.json \
  --input "We need to refactor our GraphQL API gateway to support federated schemas" \
  --text-only

bash <skill-path>/scripts/create_response.sh \
  --session-file ./out/sonnet_session.json \
  --input "Write Apollo Router configuration YAML for the subgraphs" \
  --text-only

# Multi-turn conversation by previous response ID
bash <skill-path>/scripts/create_response.sh \
  --input "How do we write integration tests for this federated setup?" \
  --previous-response-id "resp_sonnet_456" \
  --text-only

# Interactive terminal chat
bash <skill-path>/scripts/create_response.sh --chat