fix(v1): end the Bash chat program cleanly on context overflow when it cannot compact

[parkerpettit]

The bundled chat program re-raises a context-overflow 400 whenever it runs as the Bash harness, on the assumption that Bash compacts. With compaction disabled the Compactor re-raises the overflow untouched, so a Bash rollout that runs out of context exits 1 and is recorded as stop=ProviderError (and retried, if ProviderError is in the retry list) instead of ending with the transcript so far like the Null program does. The same happens with compaction enabled but no threshold, i.e. none given and none discoverable from the provider's model card: the compactor never acts in that state.

Gate the re-raise on whether the compactor can act (enabled with a threshold) instead of on args.bash. When it can, nothing changes: an overflow that reaches the loop is still a real failure, and CompactionFailed still propagates (#2584). No other harness ships this program (browser_use bundles its own).

Checked through the real rollout (interception server, subprocess runtime, scoring) against a scripted fake OpenAI server: Bash without compaction, and Bash with compaction but no discoverable window, now end ok and scored on overflow, like Null; Bash with a threshold still recovers from a mid-run overflow; a rebuilt conversation that still overflows and an unrelated 400 both still fail the rollout.

Note

Fix run_chat_loop context overflow handling based on compaction state

• Removes args.bash special-casing for API errors in the run_chat_loop conversation loop.
• Uses the compactor's enabled state and discovered threshold to determine if compaction is available.
• Ends the run with the transcript on context-overflow errors when compaction is unavailable, and propagates the error when compaction is available. Non-overflow API errors still raise.
• Behavioral Change: context-overflow errors now raise when compaction is enabled and has a threshold, instead of ending the run with the partial transcript.

^{Macroscope summarized bc4f235.}

[macroscopeapp]

Approvability

Verdict: Approved at bc4f235

Macroscope's review found this PR approvable — This one-file fix cleanly distinguishes unavailable compaction from failed compaction, preserving partial transcripts when no threshold exists while still surfacing genuine compaction failures. The change is small, localized, and does not introduce broader API or infrastructure risk.

^{You can add or adjust custom eligibility rules. Learn more.}