The degradation signature to recognize in a scenario stem: in extended sessions, the model starts giving inconsistent answers and referencing "typical patterns" in general terms, rather than the specific classes, files, or functions it actually discovered earlier in the same session. That shift — from specific to generic — is the tell that context has been lost, not that the model has stopped trying.
Five remedies, falling into two clearly separable buckets:
Bucket A — Persistence mechanisms (keep specific findings from being lost):
- Scratchpad files persisting key findings across context boundaries, so later questions in the same session can reference concrete facts instead of the model reconstructing them from a compressed memory of its own earlier work.
- Structured state exports for crash recovery — each agent exports its state to a known location; on resume, the coordinator loads a manifest and injects the relevant state into agent prompts, rather than restarting exploration from zero.
- Phase summarization — summarizing one exploration phase before spawning subagents for the next phase, and injecting that summary into their initial context, so each phase starts informed rather than blind.
Bucket B — Context load reduction (keep the main thread's own context from filling up):
- Subagent delegation for verbose investigation work ("find all test files," "trace the refund flow's dependencies") while the main agent keeps only high-level coordination in its own context — pushing the verbose part of exploration off the main thread entirely.
/compactto reduce context usage during extended sessions that have accumulated a lot of verbose discovery output, when the session needs to continue rather than restart.
The two buckets solve different halves of the same problem: Bucket A makes sure specific findings survive somewhere outside the conversation; Bucket B makes sure the main agent's own context doesn't fill up with verbose detail in the first place. A scenario can call for either or both — a session with lots of prior findings that need to survive a long conversation wants Bucket A; a session where one agent is drowning in verbose exploration output right now wants Bucket B.
The discriminator: generic answers about a codebase the agent already explored means the specifics have left context. The fix is persistence outside the conversation — a scratchpad, an export, a summary injected forward — not re-explaining or re-prompting inside the same degraded context.
Traps to recognize:
- Re-reading the codebase from scratch — expensive, and doesn't actually address why the earlier findings were lost.
- Reaching for a larger context window as the primary fix — it delays the failure signature rather than removing its cause, the same trap as in §1's summarization-loss discussion.