Current Context Policy

DuckAgent projects session history before every main model request. This keeps the prompt cache-friendly: the stable system prompt stays fixed, while dynamic context and recoverable history are appended or projected after that stable prefix.

The runtime entry is:

src/agent.rs -> projected_main_model_messages()
src/context_projection.rs -> project_main_history()
src/context_projection.rs -> ContextProjectionPolicy::guarded_mid(...)

This maps to the benchmark strategy named recoverable decay guarded mid.

Policy shape

Part	Runtime behavior
Current Agent Loop	Stays raw while the projected prompt is below the active-loop pressure threshold.
Pressure threshold	80% of prompt window for models below 200K; 85% for 200K+ prompt windows.
Current-loop projection	When pressure appears, tool results are compacted into [TOOL RESULT SUMMARY] blocks.
Completed-loop projection	Completed tool loops are projected into user messages plus one [COMPLETED AGENT LOOP SUMMARY] before model send.
Exact evidence budget	18K tokens for the active loop, 2K tokens for completed loops.
Preview budget	220 tokens for generic tool previews.
Recovery handles	File path, requested offset/limit, next offset, process id, cursor, mode, and query are preserved when available.

Tool result projection

read_file gets the richest treatment. If the file content fits inside the remaining exact-evidence budget, the projected prompt keeps exact lines. If it does not fit, the prompt keeps a compact preview plus recovery instructions:

content_recovery: call read_file with the same path and a narrower offset/limit for exact lines.

Process output keeps process identifiers, status, cursor, truncation state, a preview, and a recovery rule:

output_recovery: call process_read with process_id plus cursor/mode/search query for exact log chunks.

Other tools keep a short preview. The goal is not to hide evidence. The goal is to keep the model prompt small and cache-friendly while making exact recovery cheap and explicit.

Benchmark match

Evidence	Match
Source constructor	ContextProjectionPolicy::guarded_mid
Source threshold constants	SMALL_CONTEXT_ACTIVE_PRESSURE = 0.80, LARGE_CONTEXT_ACTIVE_PRESSURE = 0.85, LARGE_CONTEXT_THRESHOLD = 200_000
Source budgets	ACTIVE_EXACT_EVIDENCE_BUDGET_TOKENS = 18_000, COMPLETED_EXACT_EVIDENCE_BUDGET_TOKENS = 2_000
Benchmark source policy	duckagent_recoverable_decay_guarded_mid
Generated report label	duckagent_recoverable_decay_guarded_mid

The generated report label is the user-facing benchmark name. The thresholds and budgets align with the runtime implementation.

What the benchmark does not prove

The benchmark is an offline simulator. It compares policy shapes and model-pricing sensitivity, but it is not a live bill and it does not run a provider tokenizer or a real summarizer for every request. Use it to compare directionally meaningful tradeoffs:

• prompt-cache behavior;
• projected prompt size;
• recovery tokens and extra recovery tools;
• tool-output compression savings;
• model limit pressure across pricing profiles.