Summary

agent/context_compressor.py Pass 3 of _prune_old_tool_results truncates large tool_call arguments strings by byte-slicing to 200 chars and appending the literal sentinel ...[truncated]. The result is invalid JSON — mid-value, unterminated string.

Some providers (notably Anthropic via LiteLLM) re-parse function.arguments when rebuilding tool_use blocks. Once a session is compressed with any tool_call whose original args exceeded 500 chars (routine for tools like patch), the next API call fails with:

HTTP 400: litellm.BadRequestError: AnthropicException -
Failed to parse tool call arguments for tool 'patch' (Anthropic tool invoke).
Error: Unterminated string starting at: line 1 column 35 (char 34).

All 3 retries send the same corrupted messages array so they fail identically, and every subsequent user turn replays the poisoned history. The session is permanently stuck and the user sees only a canned error fallback with no way to recover (the corruption is in server-side state).

Reproduction

Any session that compresses while carrying a >500-char assistant tool_call will produce invalid-JSON arguments. On the next turn, the provider adapter (e.g. litellm's Anthropic translator) cannot re-parse the tool_call and rejects the request.

Fix

Replace the byte slice with a compact JSON-sentinel object that preserves provenance (original length + 200-char preview) while staying parseable:

{"_truncated": true, "_original_length": 1247, "_preview": "..."}

Test plan

• New unit test test_pruned_tool_call_args_remain_valid_json asserting pruned arguments parse via json.loads and carry the sentinel shape
• All existing TestTokenBudgetTailProtection tests still pass

Observed in production

Caught after a user's agent became unresponsive. A session had 11 prior patch tool_calls poisoned by this truncation after compression. Every subsequent inbound message triggered a new 3-retry failure cycle — user received 3 cryptic LiteLLM 400 error strings back-to-back.