Context in large codebases

• Context degradation shows up as inconsistency and generic descriptions replacing concrete findings.
• Scratchpad files persist key findings across context boundaries; later questions refer back to them.
• Subagents isolate verbose exploration output; the main agent keeps the high-level plan.
• Summarize one phase's findings before spawning the next phase's subagents and inject the summary into their initial context.
• Crash recovery: each agent exports structured state to a known location; the coordinator loads a manifest on resume and injects state into agent prompts.
• /compact reduces context during an extended session when discovery output has filled it.

• Hand specific questions ("find all test files", "trace the refund flow") to subagents.
• Keep a scratchpad of findings and use it when context is compacted or lost.
• Design manifests and state exports for resumable multi-agent tasks.
• Choose between /compact, a scratchpad and a fresh session with a summary.

• "A larger window prevents degradation." Noise accumulates regardless of size.
• "Saving transcripts is the same as saving state." Transcripts are mostly tool output and overflow on resume.
• "Subagents know what the main agent found." They see only what is put in their prompt.

Options that restart from scratch, re-paste earlier answers, or truncate searches. The best answer preserves specific findings in a small, durable form and keeps the noisy work out of the main context.

• 1.3Subagent invocation and context passing
• 1.7Session state, resumption and forking
• 3.4Plan mode vs direct execution
• 5.1Conversation context management