Upgraded harness

[xprilion]

This pull request introduces several improvements to the agent's context management, prompt handling, and compatibility with Anthropic's API. The main focus is on making context compaction more robust and research-friendly, improving tool output handling, and ensuring system/user message alternation is Anthropic-compliant. There are also enhancements to the system prompt to clarify persistent memory and long-running task usage.

Agent Context Management & Compaction:

• Implements a structured, four-phase context compaction in ContextManager, including pruning old tool outputs, protecting a token-budgeted tail, generating structured research summaries, and assembling compressed messages. The summary now uses a research-adapted template and supports iterative updates. [1] [2] [3]
• Adds a more accurate token estimation using tiktoken if available, falling back to the previous method if not.

Prompt & System Message Improvements:

• Updates system_prompt.yaml to inject memory, project, and knowledge context dynamically, and provides clearer instructions for using persistent memory, session search, and long-running background tasks. [1] [2]
• Per-message mode hints are now appended to user messages (instead of using system role), ensuring compatibility with Anthropic's strict alternation rules. [1] [2]

Anthropic API Compatibility & Tool Output Handling:

• Refactors message conversion for Anthropic to merge consecutive tool outputs into a single user message, and merges consecutive user messages to comply with Anthropic's user/assistant alternation.
• Sets system prompt as a typed block and adds prompt caching headers for Anthropic API calls. [1] [2]

User Guidance & Nudges:

• Appends urgent compaction and knowledge persistence hints directly to the last user message, and ensures only one hint is injected per loop iteration, prioritizing doom loop detection. [1] [2]

These changes collectively improve the agent's ability to manage long research conversations, reduce context loss, and provide a more seamless experience across different LLM providers.

---

Agent Context Management & Compaction

• Implements a four-phase context compaction: pruning old tool outputs, protecting a token-budgeted tail, generating/updating structured research summaries, and reassembling messages. Adds a research-specific summary template and iterative summary updates. [1] [2] [3]
• Improves token estimation by using tiktoken if available, for more accurate context management.

Prompt & System Message Improvements

• Enhances system_prompt.yaml to inject memory, project, and knowledge context, and provides explicit instructions for persistent memory, session search, and long-running tasks. [1] [2]
• Changes per-message mode hints to be appended to user messages, ensuring Anthropic compatibility. [1] [2]

Anthropic API Compatibility & Tool Output Handling

• Refactors Anthropic message conversion to merge tool outputs and user messages, ensuring strict user/assistant alternation.
• Updates Anthropic API calls to use typed system prompt blocks and enable prompt caching. [1] [2]

User Guidance & Nudges

• Appends compaction and knowledge nudges directly to the last user message, and ensures only one hint is injected per loop iteration, prioritizing urgent hints. [1] [2]

[sonarqubecloud]

Quality Gate failed

Failed conditions
19 Security Hotspots

See analysis details on SonarQube Cloud