Agent in long sessions emits invalid Edit/Write tool calls due to context degradation

Problem

After a long session, the agent can get stuck in a loop where it repeatedly calls

d it can no longer reconstruct the exact

d context. It also loses the precise text block it wants to replace.

This often happens when the agent is trying to update a small file like
.agents/MEMORY.md` after a long task. The failure cascade looks like:

1. Agent reads the file successfully.
2. It makes an Edit with an empty or malformed

d fails.
3. It retries multiple times, each time emitting another invalid Edit/Write call.
4. The runtime eventually interrupts the agent.

Impact

• Long-running sessions cannot reliably finish small edits.
• The user has to manually restart the session.
• The file can be left in a partially modified state (e.g., a missing markdown heading).

Suggested improvements

• Before every Edit, re-read the target file if the previous Edit failed.
• For small files or prepend operations, prefer Write with a full replacement instead of Edit.
• Consider a context-management safeguard that detects repeated invalid tool calls and forces a file re-read or session restart.

Environment

• kimi-code CLI built from main
• Agent working on
  .agents/MEMORY.md`
• Session interrupted after repeated Edit/Write failures

The report identifies the kimi-code CLI and .agents/MEMORY.md, but names no implementation file, entry point, or test. Start by reproducing a long-session failure involving Edit or Write, then trace the runtime handling of repeated invalid tool calls. Done means the recovery behavior is defined and verified for the reported failure cascade.