Context management in large codebase exploration

Signs that the context is overloaded

An extended session doesn't fail loudly. It starts giving inconsistent answers, and — the tell worth memorizing — it begins referring to "typical patterns" instead of the specific classes it discovered earlier.

That shift from the particular to the generic is the symptom. By the time you notice it, conclusions drawn in the preceding turns are suspect.

Someone who gives directions by landmark — second left after the gas station — knows the route. Someone who says it is somewhere on the north side is no longer remembering; they are reconstructing. Confidence sounds the same in both.

Scratchpad files

Have agents record key findings in a scratchpad file and reference it for subsequent questions. The file survives context boundaries that the conversation doesn't, so a specific finding from an hour ago is still a specific finding.

Delegate the verbose parts

Spawn subagents for particular questions — "find all test files", "trace the refund flow's dependencies" — while the main agent keeps high-level coordination. The exploration's verbose output stays in the subagent; the main context receives the answer.

Spawn them with the Agent tool. A subagent doesn't inherit the coordinator's conversation history, so its prompt must include everything it needs. A new agent started after a crash also begins from its prompt alone. Writing state to disk helps only if that state is explicitly loaded into the new prompt.

Between phases, write a structured handoff artifact containing exact findings, evidence locations, decisions, unresolved questions, and next steps. Inject that state into the next phase's initial context. A short narrative summary can orient the reader, but it must not replace the exact handoff fields.

/compact

/compact reclaims room in the context window during extended exploration. Use it when the window has filled with verbose discovery — and note it is lossy, which is why scratchpad files and summaries matter. Compaction is a relief valve, not a persistence strategy.

Crash recovery

For long multi-agent work, design for interruption: each agent exports its state to a known location, and the coordinator loads a manifest on resume and injects it into agent prompts.

Resuming a session is a different mechanism and doesn't cover this. It continues a conversation that survived, carrying its old tool results forward with it, and a crash is precisely the case where there is no conversation left to continue.

Without this, a crash three hours into an exploration means three hours of rediscovery.