Action Format Enforcement

Require one thought and one action per turn, repair or strip malformed replies, and stop the episode after repeated format failures.

Interactive agents that emit free-form shell or tool text need hard format rules in the harness, not only in the prompt. SWE-agent expects one DISCUSSION plus a single markdown code-block command per turn (JSON thought/action or XML delimiters are alternatives — pick one style per episode). Malformed output triggers an error template restating the exact shape; after a fixed streak of consecutive failures (three in the paper), terminate early rather than burning budget on unparseable turns.

Context hygiene: if a malformed turn is followed by a valid one, remove the malformed action/response from history so noise does not linger. Each well-formatted reply then becomes an in-context demonstration of correct interaction — useful “momentum” when the transcript is already long under collapsed observations. This is tool definition design meets agent control flow: the parser is part of the interface, and format errors are first-class observations in the ReAct loop.

---

Sources

• SWE-agent (Yang et al.), pp. 31–45