Common Context Optimization

The context is already large enough to require both observation masking and state compaction.

1. Consume the 2,000-line JSON once and extract only information needed for the active task: relevant errors, timestamps, IDs, counts, and the event sequence. Do not keep reasoning over the raw dump.
2. Replace the raw tool result with a short semantic reference, for example:

   ```text
   [Masked tool output: 2,000 JSON log lines consumed. Relevant findings: 3 errors, request ID ..., failure at ..., next event ...; full output remains available at the tool artifact if needed.]
   ```

   The output is far above the skill's masking thresholds (>50 lines, >1 KB, and >1,000 tokens), so it should be collapsed immediately after extraction. Do not mask it before it has been understood.
3. Compact the 15,000-token conversation into a rolling state rather than a dialogue transcript. Preserve the original system prompt and the last three messages uncompressed, and retain only the current goal, active task, status, current errors, key decisions, active files, and next steps. Drop chit-chat, intermediate tool calls, and assumptions that were later corrected.
4. Continue from that state using append-only turns. Avoid inserting or rewriting old messages, which can reduce prefix cache reuse.

Example state:

```yaml
Current_State:
  Goal: "<original goal>"
  Active_Task: "<current task>"
  Status: "<status/blocker>"
  Key_Decisions: ["<decision>"]
  Current_Errors: ["<error and location>"]
  Next_Steps: ["<next action>"]
```

For future turns, trigger this compaction at roughly every 10 turns or 8,000 tokens, whichever comes first. This reduces latency and protects the original intent without losing the evidence extracted from the logs.